Review

 www.advmat.de

Harnessing the Metal–Insulator Transition of VO2 in

Neuromorphic Computing
Parker Schofield, Adelaide Bradicich, Rebeca M. Gurrola, Yuwei Zhang,
Timothy D. Brown, Matt Pharr, Patrick J. Shamberger,* and Sarbajit Banerjee*

1. Introduction
Future-generation neuromorphic computing seeks to overcome the limi-
tations of von Neumann architectures by colocating logic and memory The rise of Artificial Intelligence (AI) is
functions, thereby emulating the function of neurons and synapses in the fundamentally transforming the engines
of scientific discovery, technological inno-
human brain. Despite remarkable demonstrations of high-fidelity neuronal
vation, and precision manufacturing that
emulation, the predictive design of neuromorphic circuits starting from are at the heart of modern economies.[1]
knowledge of material transformations remains challenging. VO2 is an While AI affords opportunities for accel-
attractive candidate since it manifests a near-room-temperature, discon- erating technological advancements and
tinuous, and hysteretic metal–insulator transition. The transition provides hope for new insights into intractable
a nonlinear dynamical response to input signals, as needed to construct scientific problems, major speedbumps
ahead include exorbitant computational
neuronal circuit elements. Strategies for tuning the transformation char-
costs, especially in the training of high-
acteristics of VO2 based on modification of material properties, interfacial performing models, as well as the lack of
structure, and field couplings, are discussed. Dynamical modulation of physical interpretability.[2,3] Exponential
transformation characteristics through in situ processing is discussed increases in computing are required to
as a means of imbuing synaptic function. Mechanistic understanding of handle the impending deluge of data—
projected to increase from 50 to 106 zet-
site-selective modification; external, epitaxial, and chemical strain; defect
tabytes in the next two decades—and to
dynamics; and interfacial field coupling in modifying local atomistic struc- embed physical intelligence from the
ture, the implications therein for electronic structure, and ultimately, the edge-of-the-internet to the heart of the
tuning of transformation characteristics, is emphasized. Opportunities are cloud.[4,5] Conventional computing is
highlighted for inverse design and for using design principles related to approaching fundamental limits imposed
thermodynamics and kinetics of electronic transitions learned from VO2 to by the thermodynamics of charge-transfer
mechanisms in electrostatically modulated
inform the design of new Mott materials, as well as to go beyond energy-
semiconductors and by the von Neumann
efficient computation to manifest intelligence. architectural bottleneck of shuttling data
between separate memory and processing
units.[6–9] For many decades, semicon-
ductor manufacturers kept pace with computing demands by
P. Schofield, S. Banerjee
Department of Chemistry downsizing transistors (Moore’s scaling);[10,11] however, the
Texas A&M University performance of complementary metal-oxide semiconductor
College Station, TX 77843, USA (CMOS) components is fundamentally limited by the room-
E-mail: [email protected] temperature Fermi–Dirac distribution of electron energies,
P. Schofield, A. Bradicich, R. M. Gurrola, P. J. Shamberger, S. Banerjee which, for instance, limits the steepness of the subthreshold
Department of Materials Science and Engineering
Texas A&M University
voltage swing to 60 mV per decade increase in current across
College Station, TX 77843, USA a transistor channel.[7–9] Additional limitations arise from the
E-mail: [email protected] breakdown of Dennard scaling of power density, which has the
Y. Zhang, M. Pharr consequence of sharply reducing the proportion of a silicon
Department of Mechanical Engineering chip that is switched at full frequency with each process genera-
Texas A&M University tion.[12] As a result, increasingly larger areas of a chip are left
College Station, TX 77843, USA
idle or underutilized, the so-called “dark silicon” problem.[13]
T. D. Brown
Sandia National Laboratories Indeed, with the data deluge having only just begun, Moore’s
Livermore, CA 94551, USA law scaling nearing saturation, and traditional von Neumann
architectures facing major roadblocks, computing stands at a
The ORCID identification number(s) for the author(s) of this article
can be found under https://doi.org/10.1002/adma.202205294.
watershed.
It is thus necessary to consider computing from wider mate-
DOI: 10.1002/adma.202205294 rial and architectural perspectives. In traditional von Neumann

Adv. Mater. 2023, 35, 2205294 2205294 (1 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

architectures, hardware supports a computing core consisting Namely, a candidate material must exhibit precisely tuned elec-
of a random access memory (RAM) unit, which holds data for tronic instabilities, and ideally, offer independent control over
active, ongoing processes, and a central processing unit, which 1) magnitude of conductance switching, 2) energy required
performs the arithmetic of calculations in a logic subunit in to induce switching, 3) heat dissipation, 4) switching time, 5)
addition to managing calculation instructions and data retrieval volatility of different states, and 6) the number of accessible
and recording to RAM. Much of the time and energy spent internal states. Figure 1 illustrates several of these transforma-
in computing pertains to hardware communication relays tion characteristics.
between the central processing unit and RAM.[14] This so-called
“von Neumann bottleneck” has been discussed for decades
with the most straightforward solutions being: 1) improving the 1.1. Metal–Insulator Transitions in VO2
efficiency of circuitry and circuitry components (which worked
well until the end of Moore’s and Dennard scaling) and 2) Of the candidates for neuromorphic computing applications,
adopting altogether new computing architectures. materials wherein Mott physics underpins electronic transi-
Nature presents a notable exemplar for a low-voltage, “dark,” tions afford particular promise by dint of their low entropy
highly energy economical architecture that avoids the separa- dissipation and ultrafast transformation dynamics.[7,35–37] VO2
tion of working memory and processing components. Although stands out for its distinctive metal-to-insulator transition occur-
the human brain itself comprises a multitude of subunits, it ring at 67 °C.[7,38] The electronic transition in VO2, which is
is uniquely capable of storing, processing, and transmitting exquisitely sensitive to the electronic temperature, lies in close
information within a singular fabric.[6,15–17] Neurons require proximity to a structural transformation wherein a low-tem-
much less energy to process and transmit data as compared to perature, low-symmetry monoclinic (M1) phase (space group
Si architectures as a result of their high synaptic interconnec- P21/c) undergoes displacive atomic rearrangement to yield the
tion and precisely correlated electrical spikes.[15,18] By altering high-temperature, high-symmetry tetragonal (R) phase (space
the potency of interaction between pre- and postsynaptic neu- group P42/mnm). This transition can be driven nonthermally
rons (synaptic weight), the brain essentially holds multiple by photoexcitation,[35] can be modulated to occur under dif-
states in memory, which can be conveyed across synapses by fering state parameters, and can be reversed upon cooling to
neurotransmitters.[19] As a computing design principle, this is revert the M1 state. The difference between forward and reverse
referred to as a neuromorphic computing architecture, which transition temperatures is the transition hysteresis (Figure 1),
the European Human Brain Project defines as “implementa- which is associated with energy overpotentials needed to
tion of aspects of biological neural networks as analogue or nucleate a new phase in either transition direction, and is dis-
digital copies on electronic circuits.”[20] tinctively sensitive to the microstructure and oxygen stoichiom-
The emerging discipline of neuromorphic computing seeks etry.[39–41] While a considerable literature contrasts the Peierls
to draw inspiration from processes observed in the brain, func- versus Mott origins of the transition in VO2, a recent dynamical
tionally reimagining neurons as materials colocating memory mean field theory model suggests that the transition from the
and processing capabilities without breaking communication, rutile to M1 monoclinic phase in VO2 is adiabatically linked to a
ideally with synaptic-neuronal-interconnect functional recon- Peierls-type transition but is truly a Mott transition in the pres-
figurability.[21,22] Although neuromorphic computation can be ence of intersite exchange.[36] As such, the bandgap of the M1
implemented in conventional Si circuitry using neural network phase varies as a function of the temperature. A recent neu-
algorithms featuring variable potency of interlayer weighting tron scattering study has found that the entropic driving forces
coefficients, this approach is prohibitively expensive with for the MIT in VO2 are underpinned by strongly anharmonic
regards to transistor count, time, and energy usage[7,17,23–28] phonons and not electronic contributions. The softer phonons
Many hundreds to thousands of transistors are required to sim- in the high-temperature tetragonal phase drive the increase
ulate the analog integration, thresholding, and firing processes in entropy that outcompetes orbital-driven stabilization of
of neurons that could instead be emulated by one or two non- the insulating phase with increasing temperature. Indeed,
linear dynamical circuit elements.[29–33] Thus, intense interest phonons account for ≈2/3 of the entropy increase across
has focused on the design, discovery, and utilization of novel the MIT.[42]
neuromorphic materials and device architectures such that the The MIT brings about a four orders of magnitude dis-
number, timing, magnitude, and interaction modes of electrical continuous change in conductivity along with switching of
pulses processed by neuromorphic elements encode logic and infrared transmittance and thermal conductivity.[43,44] Thus,
memory functions within a single fabric of dense, energy-effi- this transition serves as the basis for off/on (M1/R) high- and
cient computing primitives.[9,15,18] An extensive review of biolog- low-resistance states of the material. Figure 1 illustrates how
ical neuron characteristics with a focus toward their emulation transition hysteresis enables retaining off/on states when the
by complex oxide materials has recently been made available material is operated at a temperature falling within the hyster-
by Park et al.[34] In this work, we focus explicitly on connecting etic range, even after cessation of a transition-driving voltage.
material properties to neuronal and synaptic function and on This state retention serves as a basis for computational active
static/dynamical modulation of transformation characteristics memory. As will be described in subsequent sections, the MIT
of a canonical metal–insulator transition (MIT) material. To hysteresis can be modulated to be narrow or wide by various
be implemented within neuromorphic architectures, suitable processing parameters that affect defect dynamics and micro-
active materials must possess certain requisite material proper- structure, effectively enabling designs for high- or low-volatility
ties to emulate the spiking and plasticity of neuronal systems. in-memory computing.[39–41]

Adv. Mater. 2023, 35, 2205294 2205294 (2 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

exhibit MITs at the desired operating temperatures.[7] In this

article, we first examine the connection between neuromor-
phic properties and intrinsic material properties through the
lens of inverse and forward design. We next review synthetic
and processing conditions to modulate VO2 MIT characteristics
encompassing site-selective modification, point defects, strain,
particle morphology, and extended defects as a means of sys-
tematically tuning neuronal and synaptic function. The various
knobs available for tuning transformation characteristics of VO2
make it a versatile and rich palette for realizing neuromorphic
function; however, the strong coupling of spin, charge, atomic,
lattice, and orbital degrees of freedom represents a challenge
for disentangling transformation characteristics. This strong
coupling also makes these systems exceedingly difficult to treat
using density functional theory, somewhat limiting the predic-
tive power of first-principles simulations.[36,45–47]

Figure 1. Metal–insulator transition of VO2. a) The insulating M1-phase of 2. Connecting Neuronal and Synaptic Function to
VO2 reversibly transitions upon stimulation (e.g., by temperature, current, Intrinsic Material Properties
and voltage) to the conductive R-phase of VO2 with a notable hysteretic
gap between properties (i.e., signal: conductance) associated with forward A key challenge in the discovery and implementation of neu-
and reverse transitions. Several reports have noted that the electronic tran- romorphic materials is the absence of design rules that explic-
sition can be delinked from the structural transformation and engendered
itly map specific neuronal functionality (e.g., thresholding,
with minimal lattice distortion. Widening this hysteresis (dotted line vs
solid line) enables greater state stability, serving as a means of embedding spike amplification, action potentials, relaxation times) to
low- and high-volatility in-memory computing functions. b) Site-selective intrinsic material, interfacial, and circuit properties. This con-
modification such as substitutional doping on the vanadium sublattice trasts to Si CMOS, wherein rigorous analytical frameworks
of VO2 has the potential to modify a variety of metal-to-insulator proper- unambiguously map device behavior to material properties
ties including Teq., hysteresis width, sharpness of onset, or potentially through benchmarked material compositions and well-defined
even drive the system into a new MIT regime by coupling to a dynamical physical processes that can be accurately described by static
composition field such as oxygen vacancy formation or Li-ion insertion
as examples of in-material processing, thereby affording a means of syn-
functions. Neuromorphic circuit elements exhibit complex
aptic reconfigurability, which provides a counterpoint to neuronal function and highly nonlinear responses that reflect not just intrinsic
obtained with an immutable nonlinear dynamical transition. material properties but also their coupling with interfaces and
interaction with other circuit elements, heat and charge dis-
1.2. VO2 in Neuromorphic Computing sipation pathways available within device architectures, and
prior conditioning of the input signal and the active element
In practice, nonlinear dynamical behavior of VO2 and other (Figure 2). In considering how to elicit a desired set of neuro-
neuromorphic materials depend not just on intrinsic material morphic responses through systematic control of the physics
properties but also on the coupling through interfaces to the and chemistry of VO2, it is important to understand its multi­
surrounding matrix and interactions with external circuit ele- physics and multifield coupling within device architectures that
ments. Promising work by Yi et al. directed toward the design gives rise to specific aspects of neuronal and synaptic function
of neuromorphic circuits utilizing VO2 emulated distinct elec- as portrayed in Figure 2.
trical signals generated by neurons.[8] These authors replicated The ultimate criterion for the utility of a given material for
23 distinct neuronal spiking signals including all-or-nothing neuromorphic computing is the capability for desired dynamics
firing and spike-timing dependent plasticity using simple cir- that emulate action potential spiking, plasticity, and other
cuit designs—2 VO2 memristors and 4–5 passive resistor— neuromorphic behaviors.[15] This requires understanding how
capacitor elements which would have required thousands of material properties affect the potential for and qualities of
transistors in conventional CMOS circuitry with potentially resultant dynamical behavior. As depicted in Figure 2, dynam-
lower fidelity of neuronal emulation.[8] In contrast to the Pei- ical behavior depends as much on electrical coupling to external
erls insulator NbO2 first used in neuristor demonstrations,[30,33] circuit elements as on the material properties of the neuromor-
VO2 can be switched at one-sixth of the energy (<1 fJ spike–1) phic element itself.[48] Furthermore, the dynamical instabilities
and at 100× the speed.[8] associated with negative differential resistance (NDR) and self-
The MIT of VO2 occurs at temperatures slightly lower than oscillation manifest in spontaneous localization of heat and
the ≈100 °C operating temperatures typical of data servers, current density within narrow current channels.[49,50] Finally,
within which the VO2 material would be implemented. By work to quantitatively understand the design vectors to induce
comparison, the majority of materials with high-magnitude certain classes of dynamical behavior (quasi-periodic, bursting,
conductance switching capability tend to be active at tempera- chaotic) is only nascent. Nevertheless, recent work and recon-
tures several hundreds of degrees Celsius above or below room textualization is beginning to form foundational principles
temperature, with a relatively sparse palette of materials that relating material properties to instability and dynamics.[51–54]

Adv. Mater. 2023, 35, 2205294 2205294 (3 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

dT 1
= (Gel (T ) v 2 − Gth (T )(T − T0 )) (1)
dt C 
this constitutes these systems as electrothermal memris-
tors,[55,56] and establishes a feedback loop that potentially leads
to instability and dynamics.[57,58] Here, T is the temperature of
the VO2 component, t is time, C is the thermal capacitance,
ν is voltage, T0 is ambient temperature, and Gel(T) and Gth(T)
are the temperature-dependent electrical and thermal device
conductances.
For example, when under a fixed electrical bias, an individual
device is driven toward a particular steady-state temperature by
the balance of Joule heating and thermal dissipation to the envi-
ronment (Equation (1)). However, the system never remains at
its steady-state temperature for long, since thermal fluctuations
are always present.[33] An initial perturbation in temperature
changes the electrical and thermal conductance, which in turn
modifies the Joule power dissipation. This changes the magni-
tude and/or sign of subsequent net heating, leading to a new
temperature, and so the feedback loop continues (Figure 3).
The essential dynamical question is the long-term behavior
of this feedback loop: whether the cycle 1) damps itself out, 2)
maintains itself, never settling down but never growing without
bound, or 3) grows without bound. In case (1) we say the cycle

Figure 2. Multiphysics and multiscale modulation of MIT characteristics

of VO2. To create artificial neurons and synapses with tunable function-
ality from VO2, the physics and chemistry of nonlinear dynamical elec-
tronic transitions need to be understood and controlled not just at the
level of intrinsic material properties but with regards to their coupling to
external fields across decades of length and time scales. The top panel
shows desirable attributes of neuronal and synaptic function; the middle
panel illustrates multiphysics and multifield coupling used to modulate
neuromorphic response; and the bottom panel shows illustrative com-
positional, structural, and microstructural levers available to tune mate-
rial properties. While we map specific dependences, it is worth noting
that circuit-level emergent properties represent a complex convolution of Figure 3. Dynamical electro-thermal instabilities in VO2. a) Electro-
material and field coupling parameters. thermal feedback loop that can enable essential neuromorphic dynamics
in VO2. A highly temperature-dependent electrical (and thermal) conduct-
The key to these principles is a measure of nonlinearity that ance couples with Joule heating and Newton’s Law of Cooling. Feedback
determines whether the electrothermal properties of a mate- is sustained by the ΔT resulting from iv power loss that leads to large
rial are sufficiently nonlinear to sustain controlled instability changes in conductance G. b) Conceptual relationship between electro-
and oscillation. For VO2 and other thermally activated oxides, thermal nonlinearity and potential for desired dynamics. A too small non-
linearity (1) results in a rapid damping, rather than a sustained dynamic
especially NbO2, the electrical and thermal properties are highly
element; whereas a too large nonlinearity (3) is unstable, but in an unde-
temperature-dependent. Together with thermal dynamics gov- sired, uncontrollable sense. Only (2), nonlinearities that are large enough
erned by Joule power dissipation and Newton’s Law of Cooling to be measurable but not so large as to result in uncontrollable instabili-
(NLC) ties, can lead to sustained neuromorphic dynamics.

Adv. Mater. 2023, 35, 2205294 2205294 (4 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

represents dynamical stability, in the sense that perturbations ventions in materials design. However, while specific dopant/
are self-damping and the system is driven to settle at its steady strain/field-coupling influences have become apparent, for
state. Cases (2) and (3) both represent dynamical instability, instance, doping and control of oxygen stoichiometry to tune
although there is an inherent qualitative difference between the transition temperature and hysteresis width,[60,61] specific
them, with case (2) representing controllable instability and/ linkages spanning from site-selective modification to circuit-
or oscillation. From Equation (1), the dynamical properties of level properties remain to be established.
the electrical and thermal conductance must play some role in Equation (2) indicates that nonlinearity is essential for neu-
which of the cases (1), (2), or (3) the system settles in. It is this romorphic behavior. Devices without nonlinear electrothermal
role that is sought to be modulated to enable neuromorphic conductance functions cannot manifest edge-of-chaos char-
materials design. acteristics and externally controllable instability, but devices
A particularly interesting range of conditions has been that fail to be sufficiently nonlinear are also incapable of neu-
labeled edge-of-chaos[59] wherein an external stimulus can drive romorphic dynamics. As a simple example (Figure 3B), all
transformations between self-damping and self-reinforcing else being equal, VO2 with a smaller conductance contrast
regimes. The essential abstraction of the theoretical frame- and/or less sharp transformation may not be capable of neu-
work here is the partition of a dynamical system into a smaller romorphic dynamics; whereas VO2 with a larger switching
internal dynamical “core” and a dissipative external coupling ratio and sharper transformation may be capable of neuro-
field. For devices based on VO2 and similar materials, the gov- morphic dynamics, despite the same governing Mott–Peierls
erning dynamics are Newton’s Law of Cooling, which in turn physics. However, it is worth noting that it is not true that large
is explicitly dependent on the electrothermal conductance, i.e., switching ratios and sharp transformations should be pursued
the material properties (Equation (1)). A significant recent effort at any cost: electrical and thermal nonlinearities often coexist
has related the edge-of-chaos behavior to the temperature- and interfere both constructively (Γ+) and destructively (Γ−).
dependence of the electrical and thermal conductance of the For example, if the thermal nonlinearity is large but negative
material.[51] This relation is quantitatively codified in the signs and exceeds the electrical nonlinearity, there will be the poten-
of a pair of defined nonlinearity parameters,[51,57] defined as tial for instability (Γ− > 0), but the resultant instability will be
uncontrollable (Γ+ < 0), and thus not useful for neuromorphic
 Gel′ (T ) Gth′ (T )  dynamics. As such, while doping can sometimes suppress the
Γ± (T ) =  ±  (T − T0 ) ± 1 (2) conductance contrast between insulating and metallic states
 Gel (T ) Gth (T )   of VO2, provided the change in conductance is not too small
here Γ± is a generalized nonlinearity parameter, T is the tem- as to cause oscillatory dampening, a few orders of magnitude
perature of the VO2 component, T0 is ambient temperature, decrease in conductance contrast between states is still suffi-
Gel(T) and Gth(T) are respectively the temperature-dependent cient to manifest neuron-like dynamical behavior in neuromor-
electrical and thermal device conductances, and Gel′(T) and phic applications.
Gth′(T) are the total temperature derivatives of Gel(T) and Gth(T). While the framework reported here has some key limita-
When Γ− > 0, the relative increase in the feed-forward Joule tions such as a simplified view of anhysteretic behavior and
heating term with respect to temperature exceeds that of the lack of full description of temporal variations such as chaotic or
feed-back heat dissipation term. This creates the potential for bursting signals, it establishes a basis for developing a rigorous
instability. Keeping Γ+ > 0 maintains that the core remains process to map specific neuromorphic function to intrinsic
stable; it then requires external coupling for the stability to be material, structural, and interface properties. The ability to
controllably lost. This controllable stability loss due to external incorporate site-selective modifications, local structural per-
coupling appears to be key to designing a feedback loop of type turbations, and control microstructure in VO2, as well as the
(2) that cycles but does not grow without bound. The nonlin- ability to systematically tune field coupling, as illustrated in
earity parameters in terms of the electrothermal properties turn Figure 2, makes it a promising platform for connecting target
out to be a key to other relevant dynamical questions.[51] VO2 neuronal and synaptic behavior to the necessary molecular/
electrothermal memristors require an external capacitive cou- material properties.
pling to lose stability, and the size of the critical capacitance, as
well as the resultant self-oscillation frequencies are functions
of the nonlinearity parameters. In this way, the electrothermal 3. VO2 as a Building Block for Neuromorphic
properties of VO2 determine its potential for desired neuromor-
Devices
phic dynamics, and can be systematically modulated through
the levers noted in the bottom panel of Figure 2 that are dis- The choice of VO2 as a building block for manifesting dynam-
cussed in more extensive detail in subsequent sections. The ical oscillations arises from its i) existence of low-entropy
levers denoted here fundamentally modify VO2 material proper- transitions across multiple internal states with prospects for
ties, and thus MIT transformation characteristics. The compact reaching the atto-Joule and tens of ps switching required for
modeling framework described here and extensively elaborated real-time AI;[8,62,63] ii) potential for tunable multistate con-
elsewhere,[51] connects key characteristics of neuromorphic ductance switching and reconfigurability (Figure 1) to toggle
circuits such as negative differential resistance, instabilities, between neuronal (thresholding, amplification, integration)
oscillations, and bifurcations, to intrinsic material properties and synaptic (learning, memory) modes with reconfigurability
through a set of nonlinearity parameters. This approach repre- engineered by coupling to a dynamical composition field such
sents a means of linking circuit-level behavior to specific inter- as oxygen vacancy formation or proton/Li-ion insertion;[64–66]

Adv. Mater. 2023, 35, 2205294 2205294 (5 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

iii) triggering of electronic transition by light, heat, and electric neuromorphic computing beyond energy efficient computing
fields; and iv) precision control of properties from atomistic to in the design of perceptive systems.
mesoscale dimensions through numerous finely customizable
levers for independent tuning of transformation characteris-
tics, which hold promise for approaching fundamental limits 4. Doping as a Means of Controlling the Metal–
of energy and speed. Furthermore, the transformation charac-
Insulator Transition in VO2
teristics show entirely distinctive modes and strengths of cou-
pling to external composition, strain, thermal, photoexcitation, The VO2 MIT equilibrium temperature (Teq.), hysteresis values,
and electric fields, which provide opportunities of dynamical and transformation dynamics are governed by the thermo-
control of the metal–insulator transition. Figure 4 schematically dynamics of the phase transformation and kinetic attributes
illustrates several compositional, structural, and field coupling of the phase transformation such as nucleation mechanisms
levers available to tune transformation characteristics. This per- and activation barriers, respectively. Consequently, the MIT
spective outlines the assortment of material design levers avail- characteristics are tunable by modulating phase stabilities, or
able to finely tune the intrinsic MIT properties of VO2 to serve by hindering, improving, or altogether replacing the primary
as the active material of which novel neuromorphic computing nucleation mechanism. Site-selective modification of VO2
architectures may be constructed. In this article, we have through dopant incorporation provides a powerful means of
focused on mechanistic understanding of how these levers con- modulating the MIT of VO2.
trol transformation characteristics of VO2 with an emphasis on We consider first the thermodynamic relationship between
illustrative examples instead of attempting a comprehensive the insulating (M1) and the conducting (R) phases of VO2 in
account of the extensive literature. We first discuss the role of terms of how dopants amend relative phase stabilities with
site-selective modification through dopant incorporation in VO2 regards to a thermodynamic energy landscape such as illus-
and then review the particular case of thin films with regards trated in Figure 5A.
to substrate interactions. We discuss the example of electrically On a related note, we then consider the nucleation mecha-
driven MITs, which are key to enabling novel neuromorphic nisms, transformation pathways, and activation energy barriers
computing architectures and then further discuss means of between minima on the energy landscape with regards to how
imbuing nonvolatile memory through dynamical modulation they are modified by dopant inclusion. Finally, some predictive
of local structure and transformation characteristics. Finally, we guidelines for dopant selection are described with regards to
discuss photoexcitation-driven phase transformations, mixed neuromorphic function.
optical/electrical functionality, and opportunities for VO2-based

4.1. Dopant Effects on Thermodynamic Phase Stabilities

The equilibrium temperature (Teq.) of the MIT in VO2 rep-

resents the competition between stabilities of the resultant
doped phases (i.e., the M1 or the R phase) to better accommo-
date changes to the enthalpy and entropy caused by a dopant
incorporated within VO2. Doping differently modifies phase
stabilities, as well as the magnitude of structural distortion
and change in symmetry across the phase transformation.[67,68]
Figure 5C–E illustrates several illustrative cases where the lat-
tice enthalpy and/or conduction entropy contributions to free
energy are modified. Typically, changes in the thermodynamic
stability of the M1 and R phases upon doping are driven by both
a change in the relative lattice enthalpies as well as the elec-
tronic entropy. Notably, the free energy is further modified by
the strain energy and surface energy, which at the appropriate
dimensions are of comparable magnitude to the bulk free
energy.[69,70] The difference in free energy for the two phases
can be written as

∆GM1→R = (GcR − GcM1 ) + (U SE

R
− U SE
M1
) + (USR −USM1 ) (3)

where Gc is the chemical free energy, USE is the strain energy,
and US is the surface free energy.[71,72] Local field enhancements
as well as surface and strain energy differentials can counteract
the chemical free energy terms and stabilize the R phase under
Figure 4. Harnessing VO2 for neuromorphic computing by modulating
its MIT through levers: substitutional and interstitial doping, dimensions conditions of constrained equilibrium (for instance, below a
and geometry of particles, defects spanning from oxygen vacancies to specific particle size or for a certain magnitude and sign of epi-
phase boundaries, and external, epitaxial, and chemical strain. taxial strain).

Adv. Mater. 2023, 35, 2205294 2205294 (6 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

Figure 5. Dopant modulation of thermodynamic energy landscapes. a) Schematic illustration of a free energy landscape with cubic, tetragonal, and
monoclinic structures. b) Effect of dopant ion charge and size on the MIT temperatures in VO2. Dopant ions larger and higher-valent than V4+ typically
cause decreased MIT temperatures, whereas smaller, lower-valent dopants tend to increase MIT temperatures. Some dopants show a strong propen-
sity to widen the hysteresis of transition (shown in green, yellow, and red). Other dopants (Ir4+, Ru4+) stabilize alternative polymorphs of VO2. Still yet,
other dopants (interstitial boron) imbue dynamical effects on the MIT. b) Reproduced with permission.[7] Copyright 2019, Elsevier. c–e) Modification of
free energies of M1 or R phases as a function of dopant incorporation plotted as a dotted line (undoped M1 and R phases are shown as solid lines),
illustrating several possible scenarios: c) selective enthalpic stabilization of the R phase depresses the phase transition, d) entropic changes to con-
ductance entropy stabilize the M1 phase as a function of temperature, thereby elevating the phase transition, and e) entropic stabilization of the M1
phase and enthalpic stabilization of the R phase results in an overall depression of the phase transition.

Dopant/alloying elements control transformation behavior not feature an MIT near room temperature.[76] Dopants such as
by modifying the relative phase stabilities and crystallographic Al and Cr stabilize an alternative monoclinic phase (M2, space
relationships between the two phases principally through 1) group C2/m) that also features an active MIT to the metallic R
introduction of internal chemical strain, 2) influence on elec- phase.[77–79] A reversible M2→R substantially elevated transition
tronic band structure, 3) alteration of transformation pathways, has been observed at ≈134 °C for Fe-doped VO2.[80] Interstitial
and 4) introduction/modification of defect sites and density, hydrogen incorporation stabilizes two orthorhombic O1 and O2
which can interact with incipient domains and mobile inter- phases.[81]
phase boundaries. The solubility limit of a dopant is the concentration at
If the doped M1 phase is made relatively more stable than the which the stability contributed by its increased configura-
doped R phase (such as upon substitutional Ge4+ incorporation tional entropy is equal in energy to the instability contributed
on the cation lattice),[61,73] then the observed Teq. is increased through increased structural enthalpy. Thus, properties of a
from the undoped Teq., thus requiring greater energy to initiate dopant such as size and charge with respect to V4+ play impor-
the transition from M1 to the R polymorph (Figure 6F).[61] tant roles in determining solubility limits within VO2. In turn,
Conversely, if the R phase is made more stable relative to solubility limits the extent to which a dopant can modulate MIT
the M1 phase upon doping (such as upon substitutional incor- properties.
poration of W6+ on the cation lattice),[40,75] the Teq. is decreased The variable valence of vanadium and diverse connectivity
(Figure 6D). Furthermore, some dopants do not cause shifts of vanadium-centered tetrahedra, square pyramids, and octa-
in transition characteristics, but instead stabilize altogether hedra results in a “rugged” energy landscape characterized
different VO2 polymorphs. Ir-doping for example, has been by a rich diversity of metastable polymorphs.[72] The synthetic
shown to preferentially stabilize a P-phase VO2, which does design space for stabilization of VO2 is complex and limited to

Adv. Mater. 2023, 35, 2205294 2205294 (7 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

Figure 6. Dopant influences on the structure of VO2 and MIT transformation characteristics. a) Dynamical, time-dependent forward transitions are
observed in interstitially boron-doped VO2. b) Upon cooling into M1 VO2, boron atoms reside in a metastable trigonally coordinated site; over the period
of several hours, boron migrates into its most stable, tetrahedrally coordinated site. a,b) Reproduced with permission.[74] Copyright 2020, American
Chemical Society. c) Substitutional incorporation of (hexavalent) tungsten in VO2 asymmetrically depresses Teq., with further depression achieved in the
reverse (cooling) transition. d) W doping induces anisotropic stress gradients that stabilize extended M2 VO2 domains. c,d) Adapted with permission.[40]
Copyright 2018, American Chemical Society. e) Substitutionally doped germanium raises Teq. and disfavors formation of oxygen vacancies; exploiting
oxygen-rich or oxygen-poor synthesis conditions facilitates widened or narrowed hysteresis, respectively. f) Doping Ge into VO2 effectuates increased V–V
dimerization, stabilizing the M1 phase and consequently raising the Teq.. e,f) Adapted with permission.[61] Copyright 2022, Royal Society of Chemistry.

a narrow range of oxygen stoichiometries. Figure S1A of the V4+ ion generally tend to shift the Teq. up, whereas larger cat-
Supporting Information plots the Ellingham diagram of V–O ions and greater oxidation numbers generally tend to cause the
compounds, which illustrates narrow sliver of oxygen stoichio- MIT to shift to lower temperatures as shown in a collated rep-
metries where VO2 is stabilized,[82] which in turn necessitates resentation mined from the literature in Figure 5B.[7,40,61,73,89–96]
careful control of oxygen partial pressure in physical vapor Substitutional dopants have mostly been incorporated on the
deposition processes.[83,84] Similarly, solution-phase synthesis cation lattice rather than the anion sublattice, although F-doped
requires careful control of the Pourbaix diagram with regards sublattice substitutions have been reported to decrease Teq..[97,98]
to temperature, pH, and the redox potential in order to speciate By contrast, interstitial dopants reside in holes within the host
precursors to VO2 (Figure S1B, Supporting Information).[85,86] VO2 lattice. Interstitial dopants tend to be small such as H, Li,
In hydrothermal synthesis, this is commonly achieved through B, and Na, and also tend to decrease Teq..[74,87,99,100] It is typical
judicious choice of the reducing agent and pH conditions.[85] for a dopant to exclusively incorporate in a site-selective manner
The manner of a dopant’s incorporation is an important factor either substitutionally or interstitially in the manner which best
with repercussions both for configurational entropy and lattice reduces strain in the VO2 lattice, although N-doping is a noted
strain. The extent of incorporation is given by [X] = p(α, β, γ, …), exception where energetic preferences for doping on the anion
where p describes the concentration of incorporated element X, sublattice and interstitial occupancy are comparable.[101]
as a function of processing variables (α, β, γ, …). This “incor- The importance of the coordination environment sur-
poration function” has been found to be highly nonlinearly rounding a dopant atom on Teq. is readily demonstrated by
dependent on different process variables, such as temperature, interstitially B-doped VO2, which manifests a dynamical MIT
pH, precursor concentration, etc.,[38,87,88] governed in solution- with a response dependent on the duration of time for which
phase synthesis by the complex Pourbaix diagram of vanadium the material has rested in its pretransitioning state.[74] Begin-
oxides alloyed with other cations. ning with the R phase, the reverse transition to the M1 phase
The literature most often recounts cationic substitutional initially relocates the B atoms in a metastable, trigonally coordi-
doping, in which the dopant inhabits a site otherwise filled by nated site (Figure 6B).
vanadium cations of the host undoped VO2 lattice. Cationic Over several hours the B atoms eventually relax into a ther-
substitutional dopants of smaller radii or lesser charge than the modynamically preferred, tetrahedrally coordinated site. This

Adv. Mater. 2023, 35, 2205294 2205294 (8 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

is manifested by different Teq. values depending on how long of degrees.[110,111] When nucleation constraints are removed, the
the BxVO2 was allowed to relax after the initial transition—with hysteresis depends solely on the lattice incompatibility between
little time to relax (and B atoms situated in higher energy, meta- the transformed and parent phases.
stable sites), the observed Teq. is higher than when the BxVO2 We have recently demonstrated that substitutional dopant
is allowed to fully relax over several hours or days (Figure 6B). incorporation and oxygen stoichiometry can be used to
The thermally activated B migration thus provides a means of decouple modulation of transformation temperature and hys-
defining an atomic hourglass and thermometer that couples teresis (Figure 6E).[61] We found that the substitutional inclu-
with the dynamical MIT of VO2. sion of Ge increases the energetic cost to form oxygen vacan-
cies in VO2.[61] This dopant effect in tandem with preparation
in oxygen-rich synthesis conditions enables a broad hysteresis
4.2. Dopant-Modulation of Nucleation Mechanisms and extending well below the undoped VO2 transition temperature,
Transformation Kinetics despite the thermodynamic influence of Ge4+ monotonically
raising Teq..[61,73] The depletion of oxygen vacancies to serve
To launch a solid–solid symmetry-raising Martensitic trans- as nucleation sites in this system implies that a substantial
formation, a transformation dislocation needs to be nucleated degree of undercooling is needed to nucleate and propagate the
and propagated across the material.[71,102] The phase transi- reverse transition. By preparing Ge-doped VO2 in oxygen-poor
tion in VO2 is initiated by local undimerization of V–V pairs; synthesis conditions (more likely to possess oxygen vacancies),
an ensemble of dilated V–V pairs ultimately transform into a a starkly narrower hysteresis was obtained.
shear wave that propagates across the lattice near the speed of In notable cases, introduction of a dopant results in forma-
sound.[35,103] The hysteresis of the MIT results from an ener- tion of defects with longer ranges of order that may instead
getic overpotential arising from the availability (or lack thereof) serve as nucleation sites. In substitutionally tungsten-doped
of potent nucleation sites. Potent nucleation sites may come in VO2, the presence of W induces anisotropic strain sufficient to
a variety of forms including but not limited to point defects like nucleate local M2 domains (Figure 6C,D), the extended phase
dopant sites and atom vacancies or line defects such as twinned boundaries of which serve (along with twin-planes) as the pre-
planes. Consequently, defects serve as potent nucleation sites ferred nucleation sites of the forward transition.[40,112]
during the VO2 metal-to-insulator transition because less
driving force is required to nucleate the already locally unstable
defect into the new phase. 4.3. Considerations in Doping of VO2 for Neuromorphic
Although both the forward (M1→R) and reverse (R→M1) Computing
transitions occur at temperatures governed by the relative ther-
modynamic phase stabilities as altered by site-selective modi- Materials design ideas can start to be connected to the specific
fication (Figure 5), the forward transformation from a lower- needs of neuromorphic architectures. Neuromorphic com-
symmetry M1 phase to a higher symmetry R phase is readily puting elements made from VO2 will likely need to operate
nucleated at twin planes, whereas the nucleation limitations near the ≈80–100 °C operating temperatures typical of data
are more pronounced for the reverse transition. Consequently, servers. With a Teq. near 67 °C, undoped VO2 will require modi-
the effects of a dopant on the transition hysteresis can be asym- fication to raise its Teq. by several decades. Small, substitutional
metric by modulating the forward and reverse transition mono- dopants with tetravalent-or-lower oxidation states are preferable
tonically yet disproportionately as a consequence of altering to permanently increase the Teq. Cr3+ elevates the MIT to nearly
nucleation factors. In several cases, aliovalent substitution on 100 °C at sufficiently high concentrations;[113] Fe3+ doping simi-
the vanadium lattice with higher oxidation state cations strongly larly stabilizes the M2 phase and elevates the transition temper-
suppresses oxygen vacancy formation, which then results in ature as high as 134 °C as a result of segregation of iron atoms
a substantially increased nucleation barrier and the need for to V2 zigzag chains.[80] Ge4+ is a promising candidate dopant
overcooling, which is manifested in the form of increased hys- because it has been shown capable of elevating Teq. over 90 °C
teresis.[40,60,61,104] By contrast, incorporating substitutional cati- in films,[73] and its hysteresis can be modulated depending on
onic dopant ions larger in size and lower in charge than V4+ synthesis conditions.[61] The latter important factor active
(such as Sb3+) tend to result in increased oxygen vacancies.[96] memory must be colocated with processing in neuromorphic
As another illustrative example, ab initio calculations indicate computing materials, and there is need both for easily over-
bivalent Mg2+ substitutional doping induces oxygen vacancies written, as well as easily retained, memory (Figure 1). The
as a primary means of charge compensation.[105] Increasing inverse design approach outlined in Section 2 establishes a the-
defect concentration in VO2 frequently results in a narrower oretical basis for linking material properties to neuromorphic
MIT hysteresis width. Indeed, this effect is notably pronounced circuit behavior by combining compact modeling with mecha-
for increasing concentrations of doped Ti and Au.[89,106–108] nistic understanding of how dopants tune specific transforma-
Dopants are themselves point defects as a result of distortions tion characteristics[51] but needs further elaboration to obtain
induced by the impurity ion, which generate a point of local more deterministic control of circuit behavior.
instability compared to the undoped, pristine lattice.[104,109] Con- Alternative arrangements for high- and low-volatile memory
versely, the higher the crystalline quality of VO2, the wider the complements might involve selecting a pair of dopants. Gen-
hysteresis. Indeed, at reduced dimensions, where defects are erally speaking, the effect of concentration of dopants X and
readily healed through migration to surfaces, the hysteresis Y, on property M will follow some functional (nonlinear)
between heating and cooling transitions can span several tens relationships

Adv. Mater. 2023, 35, 2205294 2205294 (9 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

M = f ([ X ] / [ X 0 ]) , M = g ([Y ] / [Y0 ]) the literature that smaller substitutional cations with lower-
 (4)
or-equal valence with respect to V4+ raise MIT transition tem-
peratures, whereas the opposite effect is true for larger cationic
where [X0] and [Y0] are reference concentrations that serve to dopants of higher valence. A phenomenological Landau model
normalize the relative effectiveness of different dopants. Con- developed by Shi and Chen predicts the following general form
sidering a system with two (or more) dopants, the function for modification of the transition temperature[117]

MM = h ([ X ] / [ X 0 ] , [Y ] / [Y0 ]) (5) Teq. Θ1 (Rd3 − RV3 4+ ) / RV3 4+ + Θ2 (υd − υV 4+ ) x

 T0 → T0 − (7)
a 
represents the observed bivariate dependence of property M,
leading to a readily defined interaction term where Teq. is the undoped metal–insulator transition tem-
perature (338 K), T0 is the “Curie–Weiss temperature” corre-
η ([ X ] / [ X 0 ] , [Y ] / [Y0 ]) = h / ( f + g ) (6) sponding to structural order, Θ1 and Θ2 are coupling constants,

Rd is the radius of the dopant ion, R V 4+ represents the radius of
where η > 1 for concentrations where dopants interact coop- V4+ ions, x is the atomic fraction of the dopant with respect to
eratively, or η < 1 for where they interact destructively. A vanadium, υd is the valence of the dopant ion, υ V 4+ is the formal
similar approach could be extrapolated to multivariate doping valence of vanadium (4+), and a is a constant pertaining to sym-
scenarios. metry relations during the lattice transformation.[117,118]
Three scenarios can be envisioned. a) Dopants which tend to Dopants that provide novel nucleation sites or affect oxygen
have the same effect on a property but as a result of different vacancies tend to have profound effects on the transition hys-
mechanisms (e.g., homovalent substitutional elements which teresis. The transition hysteresis governs the on/off (R-phase/
result in lattice-strain based stabilization of a high-temperature M1-phase) state retention upon removal of the excitation. This
phase, and aliovalent equidimensional substitutional elements serves as a basis of memory within the same material to also
which principally serve to alter the electronic structure, but be used for logic operations during computation. If nucleation
with the same resulting effect). This may enable magnification constraints are alleviated, geometric compatibility of trans-
of a particular material property beyond the solubility limit of formed and parent phases dominates hysteresis width as per
a single dopant. b) Dopants which occupy distinct positions in the nonlinear theory of Martensitic transitions.[41] Identifica-
the lattice (e.g., interstitial vs substitutional), but tend to have tion of dopants that reduce the symmetry variation across the
similar functional outcomes (e.g., depressing a critical transi- phase transformation such as through control of the middle
tion temperature, Teq.). c) Dopants which compensate each eigenvalue and cofactors of the transformation stretch tensor
other in some aspect (e.g., a larger and a smaller substitutional hold promise for systematic control of hysteresis.[41,67] Con-
atom, which could cluster and result in minimal net lattice versely, approaches to imbue nonvolatile memory storage to
deformation), while still introducing a new functionality (e.g., a the MIT in VO2 are discussed in Section 6.5. Finally, thoughtful
potent defect site for nucleation, which could potentially modify selection in codopant pairs has just started to be explored but
transformation kinetics without degrading crystallinity or sub- demonstrates fine control over MIT parameters and material
stantially modifying the transition temperature). properties through balancing of lattice distortions and charge
It is likely that an optimally doped VO2 for neuromorphic compensation mechanisms.
purposes will be doped by more than one dopant species. The interplay of material processing parameters and their
Codoping offers a means of synergy in modulation of the VO2 effect on transformation characteristics is complex, and often
MIT, and indeed offers a vast design space that has only just requires careful attention to reach optimal balances between
started to be explored. In Sr-/W-substitutionally codoped VO2 multiple defect features. For example, we have recently
for example, W6+ was effective in significantly decreasing Teq., achieved a delicate balance between W-doping concentration,
whereas Sr was used to increase the material bandgap to suit particle size, and particle crystallinity (the last being related to
application needs.[114] In a Nb-/Cr-codoped VO2, the dopants the conductance differential between insulating and conductive
have antagonistic effects as envisioned in the third scenario— phases).[119] Through judicious selection of thermal annealing
the transition hysteresis is almost entirely eliminated, whereas parameters to alleviate dopant exsolvation and encapsulation
unwanted Teq. shifts are ameliorated by selecting dopants with with a passivating layer to allow annealing-induced enhance-
counteracting decreasing (Nb5+) and increasing (Cr3+) effects on ments in crystallinity, we have achieved W-alloyed VO2 with a
Teq..[107,115] A preliminary basis for charge carrier engineering depressed phase transformation temperature but large mag-
while simultaneously shifting Teq. through codoping strategies nitude of the phase transformation.[119] Modulation of VO2 by
has been demonstrated in Mg-/Fe-codoped VO2.[116] doping, stress, and other processing parameters brings critical
material properties within range of predetermined operating
conditions while simultaneously altering material performance
4.4. Toward Generalizable Design Principles for Modulating the at those operating conditions.
MIT in VO2 For application of modified VO2 materials in neuromor-
phic computing, the following conclusions can be derived.
A great diversity of elements has been incorporated in VO2, First, active systems tend to operate at elevated temperatures,
including cationic and anionic substitutional dopants in addi- so it is advisable to include a dopant that will increase the Teq.
tion to interstitial dopants. General trends emerge throughout of the MIT. Dopant solubility limits should be kept in mind,

Adv. Mater. 2023, 35, 2205294 2205294 (10 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

when possible, to ensure the VO2 system may be doped at con- freestanding particles, the transformation has been observed
centrations sufficient to bring the Teq. in range of the desired as abrupt with no phase coexistence regime,[39,124,125] con-
operating temperature. Codoping represents a potential way to sistent with a nucleation-limited transition. Additionally, par-
push above and beyond limits of singular dopants. Next, the ticle size leads to different transformation temperatures with
hysteresis of the MIT is a useful feature to exploit for neuro- larger particles having more nucleation sites and thus faster
morphic materials so as to provide a selection of volatile or less- switching.[104,119,126] By contrast, mechanically constrained VO2
volatile memory capabilities. Finally, although codoping in VO2 shows rich phase coexistence characteristics, resulting from
remains largely underexplored, combining multiple, disparate local thermoelastic equilibrium (Figure 7). For instance, for-
dopant effects enables materials design and MIT engineering mation of periodic and dynamically tunable nanodomains in
possibilities extending beyond the scope of individually doped constrained nanobeams has been attributed to the competi-
systems. tion between elastic strain energy (from lattice mismatch) and
domain wall interfacial energies.[124,125,127] Nanometer-sized
domains are similarly stabilized in thin films but tend to be dis-
5. Thin Films of VO2 and Their Potential for ordered without long-range periodic organization.[128]
Applications in Neuromorphic Computing
Thin films of VO2 are of particular relevance due to the scal- 5.1. Lattice Strain in VO2 Nanobeams and Thin Films
ability of fabrication processes based on lithography tech-
niques and compatibility with foundry capabilities. Several Applying uniaxial tensile stress along the cR-axis modifies the
reviews have examined the deposition of high-quality VO2 bonds lengths of V–V bonds and can affect the Teq. by impeding
thin films[120,121] and the resulting processing–microstructure– V–V dimerization. This is perhaps most clearly discernible
property relationships.[122,123] Here, we will focus instead on upon application of external strain to nanobeams along specific
reviewing methods for controlling the MIT of VO2 thin films crystallographic axes. Experiments examining the application
using epitaxial strain, use of dopant incorporation, and the of external strains have revealed the emergence of metastable
use of high-energy irradiation to control defect populations, M2 and T phases, which are stabilized under anisotropic strain
and as well as the intersections between these areas (Figure 4). (Figure 7A). These polymorphs exhibit parallel transformation
These approaches are keys to developing systems with deter- pathways to the rutile phase upon heating. Park et al. strained
ministic control over transformation characteristics, as well as VO2 nanobeams to construct a stress–temperature phase dia-
in realizing systems with long operational lifetimes that are gram and found a triple point at which multiple phases are
required in computing applications (e.g., endurance beyond 109 degenerate (Figure 7A).[129] Cao et al. found that applying strain
cycles). Achieving these goals requires establishing the intrinsic to clamped nanobeams induces phase inhomogeneity; peri-
mechanical properties of the thin films, the stress evolution odic triangular domains are formed along the length of the
behavior of the films during phase transformations, and design nanobeams to relax strain and minimize interfacial energies
strategies to design damage-free operation of VO2-based neuro- (Figure 7B).[124] The distinctive strain sensitivity of VO2 is par-
morphic device elements. ticularly strong along the cR-axis, which controls the closing
MIT behavior of VO2 thin films is tied to film geometry of the bandgap, as has been seen in cluster-dynamic mean
and the corresponding boundary conditions. In undoped field theory calculations.[130] A detailed perspective of DFT and

Figure 7. Strain modulation of the VO2 phase diagram. a) Strain manifests in the VO2 system by stabilizing one phase relative to each other. b,c) Phase
coexistence resulting from local thermoelastic equilibrium. a) Stress–temperature phase diagram of VO2 resulting from tensile strain. Here Tc indicates
unstressed Teq., Ptr, and Ttr indicate the pressure and temperature associated with the triple-point of M1, M2, and R-phases of VO2, and T indicates
the triclinic T-phase VO2 structure. b) Optical images of a VO2 nanobeam heated through the transformation during application of external bending
strain, resulting in coexistence of domains in local thermoelastic equilibrium. c) Textured microstructure comprising R- and M1 phases that results
from transformation of a single-crystal VO2 nanobeam that is elastically coupled to the substrate. a) Reproduced with permission.[129] Copyright 2013,
Springer Nature. b) Reproduced with permission.[124] Copyright 2019, Springer Nature. c) Reproduced with permission.[79] Copyright 2010, American
Chemical Society.

Adv. Mater. 2023, 35, 2205294 2205294 (11 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

beyond-DFT calculation methods used to identify and investi- r-cuts of Al2O3,[135,137,138] various cuts of (TiO2)R,[131,136,139–142] and
gate electron-correlated metal oxides for neuromorphic com- implementation of buffer layers such as (TiO2)R, Al2O3, NiO,
puting is available elsewhere.[34] and SnO2.[143–147] A collated summary of the role of epitaxial
In contrast to nanobeams, stresses in epitaxial films are mismatch in modulating the MIT in VO2 excerpting a substan-
typically biaxial in nature, as they originate from in-plane lat- tial span of the relevant literature is provided in Figure 8b. Teq.
tice mismatch between the film and the underlying substrate. for VO2 scales linearly with the strain along the cR-axis on TiO2
In films under biaxial stress, the stress–temperature phase substrates. With decreasing thickness, indicated by various
diagram is modified because of the lattice mismatch, which shades of purple in Figure 8, the interface between the film and
depends on the underlying substrate. For VO2 on TiO2 (001), the substrate becomes more relevant as strain is not as relaxed.
both the aR-and bR-axes are equally strained; tension along the In TiO2 (001), with a layer of ≈7 nm of VO2, it has been docu-
cR-axis raises the temperature of the MIT, as observed in phase mented that the transformation does not occur; even at high
diagrams constructed by Park et al.[129] and experiments by temperatures, the film remains insulating as a result of inter-
Muraoka and Hiroi.[131] Hard X-ray photoemission spectroscopy diffusion of titanium atoms, which gives rise to the formation
studies of buried interfaces of VO2 thin films on TiO2 (100) con- of a new (V,Ti)O2 solid solution.[70,148]
firm that strain along the cR-axis stabilizes the M2 phase but at Related to these observations, the competition between the
a different temperature than the proposed phase diagram as a domain interfacial energy and the elastic lattice misfit energy
result of strain-induced increased electron correlation, which in has received considerable attention since it governs the dimen-
essence triggers an orbital-selective Mott transition.[132] Aetukuri sions and periodicity of phase separation (Figure 9).[136,149–152]
et al. used a buffer layer of RuO2 with varying thicknesses to A bulk single crystal of VO2 displayed a sharp metal–insulator
apply tension along the cR-axis and in-plane compression along interface, whereas a polycrystalline VO2/r-sapphire system
the aR-axis, and found that the MIT shift is correlated with the showed formation of a complex percolative network of domains
d|| orbital occupancy of the metallic R state.[133] Selective modu- across the phase transformation; the greater phase hetero-
lation of the V–V bond length through anisotropic strain allows geneity of the latter stems from several in-plane orientations
for continuous tuning of the orbital filling of V 3d hybrid states (aM-axis) existing in polycrystalline systems as compared to
oriented along the cR-axis, enabling precise modulation of Teq.. the singular orientation of single-crystalline VO2 films.[150,151]
M2 and T polymorphs are stabilized under anisotropic strain Consistent with this idea, polycrystalline VO2 thin films grown
with partial undimerization of V–V chains, thus serving as a on c-axis sapphire transform to the R-phase through a series
means of strain minimization. Unlike in nanobeams, it has of independent nucleation events; the distribution of local
been found that biaxial strain can lead to phase coexistence in strains is accumulated at the domain boundaries of VO2 upon
a more random and percolative manner for epitaxial VO2 thin thermal cycling and brings about a monotonic increase of the
films on Al2O3 ( 1 012), TiO2 (110), and Al2O3(0001).[134–136] transition temperature and decrease of the hysteresis.[135] VO2/
Applied strain through epitaxial mismatch has been real- TiO2 (110)R yields a stripe-like pattern of VO2 polymorphs as
ized with several different substrates such as the c-, m-, and a result of lattice mismatch along [110]R and [001]R (Figure 9);

Figure 8. Substrate effects on MIT characteristics of VO2 thin films. Thin films grown on rigid solid substrates invariably result in elastic interactions
between the film and the underlying substrate. When there is an epitaxial crystallographic relationship between the film and the substrate, this result
in additional anisotropic lattice strain which can stabilize one phase with respective to the other, as demonstrated by a) thermal transformation curves
on different epitaxially grown films.[131,135–147] b) MIT in VO2 films grown on different substrates are impacted by 1) the crystal structure of the underlying
substrate, 2) the thickness of the overlying VO2 film, and 3) the existence of a buffer layer.

Adv. Mater. 2023, 35, 2205294 2205294 (12 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

Figure 9. Imaging the MIT in VO2 thin films. a) Near-field IR images showing the three stages of the MIT. b) Anisotropy along the [1 10]R and [001]R
directions using temperature-dependent THz conductivity (σTHz). c) Infrared near field (S3) and AFM images of a VO2 film at different temperatures.
Adapted with permission.[136] Copyright 2013, American Physical Society.

the periodic organization of the phase domains reflects the bal- 5.3. The Intersection of Strain and Chemical Alloying
ance between minimization of lateral lattice coherency strains
and interfacial lattice mismatch.[136,151] By contrast, highly ori- Incorporation of dopants in VO2 thin films clearly modifies the
ented VO2/TiO2 (001)R with the cR-axis as the out-of-plane direc- MIT; however, differences in the effects of site-selective modi-
tion shows a gradual and continuous phase transformation fication between epitaxial and doped films are less explored. A
pattern.[150] question of particular importance is the extent to which strain
from dopants (chemical strain or pressure) couples with epi-
taxial strain derived from lattice mismatch with the substrate?
5.2. Alloying of VO2 Thin Films Epitaxial V0.97W0.03O2 deposited on a-plane (110) sapphire had
a reduction in Teq. of 23 °C at%−1 W−1 and exhibited an MIT
As discussed in preceding sections, dopants induce local struc- with decreased sharpness.[166] Use of reciprocal space mapping
tural distortions, trigger electron or hole injection, and modify in a tungsten-doped VO2/TiO2 (001) system confirmed the ten-
defect distributions. Successful early attempts at thin film sion on the c-axis arising from the larger size of tungsten.[167]
doping included incorporation of tungsten and molybdenum Furthermore, the concentration of tungsten (x in WxV1-xO2)
with depressed Teq. values of 40 and 55 °C, respectively.[153] Later systematically reduces the Teq. at concentrations x > 0.07.
works with the same dopants established a linear relationship Concentrations between 0.07 ≤ x ≤ 0.095 undergo a dramatic
between the Teq. and x ≤ 0.020 in V1-x(Mo/W)xO2.[154,155] By con- decrease in resistivity with slight upturns, which indicates the
trast, doping of titanium in amounts from 5 to 20 at% leads to formation of a metallic matrix with the presence of insulating
a smeared transformation ranging from 75 to 80 °C.[106] Dopant puddles. Further increase in dopant concentration beyond this
incorporation has been achieved through kinetic control using point leads to an increase of Teq. and a smearing of the trans-
methods such as sputtering, sol–gel processing, or chemical formation at points x > 0.20. The shift to higher temperatures
vapor deposition, wherein the kinetics of crystallization exceed with doping concentrations and its relation to strain coupling
kinetics of dopant diffusion and exsolvation.[119] Alternatively, a remains an opportunity for future investigations. Doping with
few studies have implemented postsynthetic dopant incorpora- Cr and Nb in VO2 on Al2O3 (0001) results in the reduction of
tion through methods such as ion implantation.[156] Substitu- the thermal coefficient of resistance and the Teq. through lattice
tional aliovalent dopants such as W6+, Mo5+,[154,157] and Nb5+,[158] deformation. The full scope of correlations between substrate
with higher valence and larger radii than tetravalent vana- strain and dopant inclusion remains to be mapped. Modulation
dium shift Teq. to lower temperatures. Tungsten incorporation of fundamental VO2 properties such as magnitude of switching
brings about a depression of Teq. at ≈43–55 °C at%−1 W−1.[158–160] and ΔTeq. needs to be investigated for alloyed VO2 films as a
Smaller ions that induce a relatively modest lattice distortion function of substrate coupling.
such as F− incorporated on the anion lattice and H+ incorpo-
rated in interstitial sites tend to have a less pronounced effect
on the transition temperature.[161,162] Substitutional incorpora- 5.4. Defect Incorporation through Irradiation Exposure
tion of formally tetravalent and lower valence cations such as
Ge4+,[73,163] Ti4+,[106,164] Al3+,[165] and Cr3+ tend to increase Teq.; a Particulate and electromagnetic irradiation have been used to
Teq. of ≈100 °C has been observed for V0.8Cr0.2O2 prepared by modulate the structure and MIT characteristics of VO2. Elec-
pulsed laser deposition on Al2O3 (0001).[113] tromagnetic irradiation generally changes carrier density in

Adv. Mater. 2023, 35, 2205294 2205294 (13 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

VO2 and can strongly perturb the electronic structure. X-ray based on cantilever resonant frequency measurements.[177]
irradiation systematically increases the electrical conductivity Through nanoindentation at room temperature, Jin et al.
of the M1 phase of VO2 with increasing dose, while modifying reported the “composite” modulus of polycrystalline VO2 thin
the electrical conductivity of the R phase in a somewhat sto- films on silicon as 140–170 GPa.[176] With the same technique,
chastic manner.[168] Upon prolonged X-ray irradiation, Teq. has the authors also reported the “composite” modulus of epitaxial
been found to first increase and then decrease with increasing VO2 films on sapphire as 240–260 GPa. The “composite” mod-
dose, which has been attributed to the irradiation mediating ulus in these studies includes contributions from both the film
redox reactions on the vanadium sublattice, which results in of interest (VO2) and the substrate (Si or sapphire). To remove
self-doping.[169] the influence of the substrate and thus measure the intrinsic
Particulate irradiation can be used to increase defect den- modulus of VO2, Zhang et al. utilized the Hay–Crawford thin
sity and modify defect dynamics in VO2, resulting in substan- film model in conjunction with experimental studies, and
tial alteration of thermodynamic stabilities and transformation reported the elastic modulus of VO2-M1 and VO2-R sputtered
pathways. Neutron irradiation has been shown to decrease the thin films as 224 ± 20.0 GPa at 25 °C and 210 ± 15.5 GPa at
electrical conductivity of the M1 phase of VO2, which is attrib- 85 °C.[180] In terms of nanowires, Minor et al. estimated the
uted to irradiation-induced creation of defects and expansion of elastic moduli of VO2-M1 and M2 nanowires to be 128 ± 10 and
the lattice.[170] In epitaxial VO2 thin films, proton irradiation sys- 156 ± 10 GPa through tensile tests inside a TEM.[13]
tematically increases the electrical conductivity of the M1 phase It is important to determine stress levels that develop in
with increasing ion fluence while decreasing Teq.[171] Alpha- physically constrained environments during MITs, so as to
particle (He2+) irradiation has been shown to decrease Teq. in guide the design of practical neuromorphic devices that avert
freestanding VO2 nanobeams, while increasing the tempera- damage during operation. Zhang et al. utilized a multibeam
ture of the MIT in clamped nanobeams.[39] This phenomenon optical stress sensor to assess the influence of crystal orienta-
has been attributed to the creation of vacancy and interstitial tion on the stress evolution of VO2 thin films across repeated
defects during irradiation. structural phase transformations. Textured polycrystalline VO2
Much remains to be explored about how combinations of films deposited on Si substrates, as well as epitaxial VO2 films
irradiation exposure and chemical alloying as well as irradiation deposited on sapphire (0001) substrates, develop large tensile
exposure and strain modify MIT characteristics. Systematic stresses when heating through the MIT, whereas epitaxial VO2
exposure of VO2 to electromagnetic and particulate irradia- films deposited on TiO2 (001) substrates develop large compres-
tion under different boundary conditions holds promise for sive stresses in heating through the MIT. This seeming disparity
advancing fundamental understanding of defect formation, is attributable to the highly anisotropic deformation that occurs
defect diffusion, and interactions of point defects with each during the phase transformation. Mapping atom positions
other and domain boundaries. At the systems level, memris- during the phase transformation by reconstructing the trans-
tive devices have the potential for greater radiation hardness formation tensor corroborates that the strain levels and signs
compared to Si devices and can potentially encode irradiation- (compression vs tension) are consistent with the experimental
induced reconfigurability as an intrinsic functional element.[172] observations. These studies highlight the importance of the
crystal orientation of VO2 films on the stress levels developed
during thermal cycling, which governs the probability for failure
5.5. Mechanical Resilience of VO2 Thin Films: Elastic Modulus, (e.g., fracture primarily as result of tensile stresses). In another
Dynamical Evolution of Stress across the MIT, and Fracture study, Viswanath et al. utilized a multibeam optical stress sensor
Behavior to monitor the stress evolution of polycrystalline VO2 thin films
deposited on Ge(100) substrates during thermal cycling.[181]
Given the substantial change in crystal symmetry and lattice They found reversible stress evolution over 100 thermal cycles
volume across the M1–R structural phase transformation and through the MIT in their films.[181] While these results are prom-
the anisotropic strains engendered in the material, VO2 films ising and stress is a useful tunable parameter for modulating
often suffer severe mechanical damage upon thermal cyc phase transformation characteristics, active neuromorphic ele-
ling.[102,140,141,150,152,173,174] In this section, we review the mechan- ments will need to survive many orders of magnitude more
ical properties and deformation behavior of VO2 thin films. on/off cycles. During such long-term cycling, VO2 films may
While mechanical stresses can be used as a tuning parameter undergo damage from effects of fatigue. Therefore, studying the
for modulating the MIT characteristics of VO2 (Figure 4),[129,175] fatigue behavior of VO2 thin films will be critical in designing
thin films are prone to damage if sufficient stresses accumulate mechanically robust VO2 neuromorphic elements.
during thermal cycling. To accurately determine stress values at In terms of fracture behavior, several groups have observed
a certain level of strain, the elastic modulus of VO2 must first cracking in VO2 thin films.[102,140,141,152,174,182–184] Nagashima
be determined.[176–179] Singh and Viswanath surveyed the litera- et al. studied epitaxial VO2 (402)M/TiO2 (001) thin films of
ture and found the elastic modulus of the insulating M1 phase various thickness ranging from 10 to 30 nm.[140] Their XRD
to range from 100 to 308 GPa based on experiments and first- studies indicated that the crystal structure of the thin films did
principles studies.[179] The substantial spread of values derives not change significantly with varying film thickness. However,
from differences in geometry, crystal orientation, testing they found that electrical transport properties changed from a
conditions, and fabrication conditions, as well as the poten- sudden jump in electrical resistivity with temperature to a more
tial role of experimental artifacts. In terms of thin films, Rúa complicated multistep process during the MIT with increasing
et al. reported the elastic modulus of thin films of VO2-M1 as film thickness. Likewise, the on/off ratio of electrical resistivity
156 ± 7.5 GPa and thin films of Cr-doped VO2-M2 as 102 ± 3 GPa during the phase transformation decreased with increasing

Adv. Mater. 2023, 35, 2205294 2205294 (14 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

film thickness. The authors argued that the disparities between of by indirect thermal activation. Many prior studies have
the observed crystal structure (which remained relatively focused on deciphering the origins of the electrically driven
similar with varying film thickness) and transport properties transition, which can generally be differentiated into two broad
(which changed with varying film thickness) stemmed from the mechanisms: an electronic-structure-modulated MIT versus a
presence of nanoscale cracks, which grew when the film thick- Joule-heating-induced MIT. The notion of a carrier-injection-
ness was above 15 nm.[140] Paik et al. reported a similar trend mediated electronic transition is attractive for allowing for
between the on/off ratio of electrical resistivity and critical frac- delinking of the electronic phase transition from the structural
ture thickness of epitaxial VO2 (402)M/TiO2 (001) thin films.[141] phase transformation, which can enable computing approaches
When sufficient lattice strain is induced in VO2/TiO2 (001), that approach the limits of energy dissipation.[185,186] Proposed
VO2 has been found to stabilize as the metallic phase (VO2-R) mechanisms for this model include carrier injection, photoexci-
at room temperature (and even below).[140] However, under tation, or a percolative-avalanche model.[185–188] However, some
these same conditions, Rodríguez et al. found that VO2 in of these models have been critiqued for insufficiently consid-
regions near a crack existed as the insulating M1 phase, whereas ering interfacial thermal resistance or heterogeneous transport.
regions away from the crack existed as the R-phase.[174] This pat- Examples of experimental results include characterization of
tern of phase separation can be attributed to the fact that epi- the electrically driven MIT in which it is found that the elec-
taxial lattice strain is relaxed near the cracked region when the tric field needed to induce the phase transition is variable
film thickness is large enough to induce fracture of the films. depending on the type of metal used in the electrodes, as well
As such, it is clear that there is strong interconnectedness and a as Raman spectroscopy results showing that the current- or
delicate balance between film geometry (orientation and thick- temperature-dependent spectra for the electrically driven MIT
ness), fracture, and phase stability (e.g., near/away from frac- do not exactly correspond to the thermally driven MIT.[185,188]
ture regions in which strain is relieved), which correspondingly With respect to previous electrothermal models of a VO2 device,
influence the MIT behavior. localized current density and the formation of R phase chan-
Lattice strain, chemical alloying, and irradiation to induce nels are not seen.[186,187] An alternative set of mechanisms used
defects are important tunable parameters for modulating the to explain the behavior of the electrically driven MIT invokes
MIT characteristics of VO2, and thus neuromorphic function an electrothermal MIT that is driven by Joule heating.[189–194]
(Figure 4). Lattice strain, while effective in depression/eleva- Indeed, more recent works attests to Joule heating as the pri-
tion of the Teq., can be relaxed through deformations and is mary mechanism[189–194] in light of linear relationships meas-
not always equally dispersed throughout the film. Advances in ured between power dissipated by the device at its critical
chemical alloying have resulted in identifying the most effec- switching voltage, which is proportional to Vcrit.,[2] and the
tive dopants at shifting the MIT but this method introduces ambient temperature, from which it is inferred that the tem-
large amounts of spatially varying strain on the lattice, which perature at Vcrit. corresponds to Teq.[191,192,195,196] This has further
can yield complex stress gradients across phase transforma- been quantitatively confirmed by direct measurements of the
tions. Finally, irradiation can not only shift the transition but temperature of the material during the electrically driven phase
also increase the conductivity through mediation of redox transition, which demonstrates the formation of localized hot
reactions that introduce charge carriers. However, the effects filaments.[194,197]
of irradiation are less predictable and damaging at high ener- In a two- or three-terminal VO2 device, the MIT can be trig-
gies. In terms of mechanical integrity, we recommend several gered by applying either a voltage or current to indirectly heat
strategies to move toward damage-free operation of VO2-based a volume of VO2; the two stimuli result in markedly different
devices. First, decreasing the thickness of the film is an effec- transformation behavior. The temperature of the device is con-
tive technique to reduce the crack driving force and thus avert trolled by the energy balance between the generated heat and
fracture. Uncracked films have shown better performance the heat dissipated away from the device (see Equation (1)).
characteristics, e.g., on/off ratio. Next, adding an appropriately Under an applied voltage, the phase transformation occurs
designed buffer layer between the film and substrate may effec- when the temperature reaches Teq., corresponding to the
tively decrease the lattice mismatch between the VO2 film and device’s critical voltage (Vcrit.). This coincides with an abrupt
substrate, thereby reducing tensile stresses generated during increase in the heat generated by the device, which causes a
operation and thus the propensity for fracture. Finally, imple- sharp drop of its resistance (Figure 10). This causes a feed-for-
menting substrates that possess similar coefficients of thermal ward effect on the temperature of the device, which drives the
expansion to the VO2 films can minimize thermally induced MIT to some state of completion. When instead a current is
stresses. Many of these films are fabricated at high tempera- applied to the device, the MIT’s accompanying drop in resist-
ture; matching the coefficients of thermal expansion of the sub- ance decreases or stabilizes the heat generated by the device,
strate to the film minimizes thermal stresses that arise during allowing the phase change to occur over some range of current
heating/cooling. values in a feed-back effect (Figure 10).

6. The Electrically Driven MIT and Electrothermal 6.1. Spatial Inhomogeneities in VO2 under an Electrical Bias
Couplings in VO2 Devices
The I–V response of VO2 devices is locally active within its
For efficient and fast computational applications, the MIT in region of NDR, meaning that in this region of I–V space, the
VO2 will need to be driven by an applied electrical bias instead VO2 element itself controls either the net current through

Adv. Mater. 2023, 35, 2205294 2205294 (15 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

Figure 10. Voltage- and current-driven MIT in VO2. a) Current–voltage response of voltage- (black) and current-driven (red) MITs in a VO2 thin film
device fabricated from a 100 nm polycrystalline VO2 thin film deposited using reactive magnetron sputtering. The critical switching voltage (Vcrit.) and
critical switching current (Icrit.) are labeled. While the voltage-driven MIT occurs abruptly, the current-driven MIT occurs over a range of current values.
b) This phenomenon is seen also in the measured resistance of the device through the phase change, as the resistance value of the device changes
abruptly in the voltage-driven MIT but gradually over a range of values in the current-driven MIT.

the circuit (when the external source is an applied voltage) or ence during the electrically driven transition has interesting
the voltage across each circuit component (when the external implications with respect to tailoring the device behavior toward
source is an applied current).[8] NDR refers to an electrical specific desired properties. For instance, with control over the
resistance which decreases as a function of an increasing total volume of VO2 undergoing the MIT, the conductance dif-
applied electrical bias ferentials between heterogeneous phase domains become tun-
able with domain periodicity, affording finer control over the
dI (8) delay time for switching the device.
<0
dV  Altering the size of the conductive channel is crucial to the
performance of the device. It has been demonstrated that after
NDR can be found in materials that form higher current the appearance of the conductive channel, additional applica-
density channels relative to the rest of the material under the tion of electrical stimuli tends to widen the channel.[197,201]
application of an electrical bias.[50] The formation of conduc- Other studies have focused on tuning the size of the conductive
tive channels of R phase VO2 has been demonstrated through channel using different series resistors, thereby modulating the
different methods (Figure 11).[190,195,197–199] These filaments can ratio of M1 to R phase present in the device based on the ability
be observed optically as a result of the difference in reflectivity to control the dimensions of the ON-state channel.[192,200] The
between the two phases. Spatial mapping of X-ray signals also size of the channel is also dependent on the device geometry, as
reveals localized formation of the R phase.[200] Phase coexist- there is a linear relationship between device length and channel

Figure 11. Inhomogeneity in the MIT. Several methods have been used to characterize the emergence of spatial inhomogeneities across the electrically
driven MIT. a) Optical images show that the width of a conduction channel can be tuned by the magnitude of the external load resistor. Reproduced
with permission.[196] Copyright 2014, AIP Publishing. b) Other optical studies focus on the delay time of the onset of the filament formation. Reproduced
with permission.[190] Copyright 2013, AIP Publishing. c) The phase of the conductive filament has been confirmed to be rutile VO2 by spatial mapping
of X-ray analysis. Reproduced with permission.[200] Copyright 2013, AIP Publishing.

Adv. Mater. 2023, 35, 2205294 2205294 (16 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

width, but relatively little correlation between device width and active element is a major determining factor in its response to
channel width.[196,202] Another study shows that whereas for electrical stimuli. Nanobeam prototype devices are of interest
larger devices (L > 10 µm), the channel appears at the same in fundamental materials research as they behave like quasi-1D
position each cycle, for smaller devices (L ≤ 10 µm), there is structures that are confined along two out of three dimensions
some stochasticity in the position of the channel from one cycle and can be strained along specific axes.[125,195,207] For thermal
to another.[196] triggering of phase transitions, this is manifested in abrupt,
single-step MIT behavior for undoped and unstrained VO2
single crystal nanostructures, which are further characterized by
6.2. Energy Efficiency and Switching Speed of VO2 Devices a large RON/ROFF ratio between the two resistance states.[104,125]
When the transition is electrically induced in unstrained single-
Improved control of VO2 MIT transformation characteristics crystal VO2 particles, a striped pattern of M1 and R phases
is critical to the design of neuromorphic device elements that emerges instead of the channel formation observed in thin film
overcome the prohibitive bottlenecks faced by digital AI at all devices, and reflects the balance between interfacial coherency
levels of the stack. To realize the full promise of neuromorphic strains and beam deformation.[195,208] This results in a stepwise
computing, new algorithms and computing architectures will sequence of resistance states, which stands in contrast with the
also need to be codesigned alongside circuits that exploit the single-step phase transition behavior observed in the thermally
tunable characteristics of the VO2 MIT. A deeper discussion driven MIT.[60]
of the fundamental approaches behind historical and current In VO2 thin films, a comparison can be made between the
artificial neural network models and modularized neuronal properties of polycrystalline thin films and epitaxial or single-
and synaptic hardware development has recently appeared else- crystal thin films because properties such as grain size, defect
where.[34] To compute the energy advantages of individual stack density, and crystal orientation of the films affect their ther-
components, a rigorous evaluation of switching energy and mally driven MIT properties such as the ROFF/RON ratio, Teq.,
speed is required, albeit such projections are most useful when and hysteresis.[123,193,194] However, it is challenging to make a
a functional chip-scale computing advantage can be evaluated. direct comparison between polycrystalline and epitaxial thin
Table 1 presents a compendium of switching energies and film device responses to electrical stimuli, since it can be dif-
times found and inferred from the literature for VO2 devices ficult to disentangle the role of film texture from extrinsic
as a function of device dimensions. In general, the switching parameters such as device dimensions. While Vcrit., Icrit., and
energy and speed of VO2 compares favorably with that of the Teq. can be compared directly, it is unclear how other character-
Peierls’ insulator NbO2, whose much higher MIT tempera- istics, such as hysteresis, directional electrical conductivity, and
ture requires 6× the energy and 100× the switching time.[8] In single-phase electrical conductivities will affect the electrically
order to develop accurate stack-level projections, the energy driven MIT, in addition to the effects caused by differences in
and time scaling of electrical switching in VO2 must be better thermal properties between a deposited VO2 layer and its sub-
understood in terms of the various parameters that affect these strate. Additionally, the role of crystallinity and texture of VO2
outputs. These parameters include external circuit elements, on formation of conductive channels remains unexplored.
ambient temperature of the device, device geometry, and device
thermal and electrical properties as well as the MIT and mate-
rial properties of VO2 (which are tunable based on the para­ 6.4. Effect of Site-Selective Modification on VO2 Device Behavior
meters shown in Figure 4).
The incorporation of dopants can alter VO2 device response to
electrical stimuli. Tungsten-doped VO2 nanobeams and nano-
6.3. Electrically Driven MIT in Undoped VO2 Particles and films exhibit decreasing Vcrit. as a function of increasing W
Thin Films concentration, which is correlated to a lower Teq. as compared
to undoped VO2.[209,210] In contrast to the linear relationship
Different aspects of the electrically driven MIT can be tuned between V2crit. and temperature in undoped VO2, the critical
to modulate VO2 device performance. The morphology of the switching voltage for the MIT in W-doped VO2 follows an

Table 1. Examples of energies and times to switch VO2 devices. L, W, and t refer to length, width, and thickness, respectively, where L is the distance
between the two electrodes. Additional factors that would affect the switching energy and time, but are not listed in this table, include the ambient
temperature of the device, device properties such as interfacial thermal resistances, and the variable interfacial and material properties of the VO2
films used to make these devices. Where the energy of the device was not found in the text, the power to switch the device was multiplied by the
switching time to provide a value for the switching energy.

Device geometry L W t Switching time Switching energy Refs.

Planar 3 µm 30 µm 9 ns 0.35 nJ [203]
Planar 0.2 µm 2 µm 50 nm 100 ns 8.58 nJ [193]
Through-plane d = 0.2–1 µm d = 0.2–1 µm 400 nm 1.9 ns 0.11 nJ [204]
Nanogap 125 nm 5 ns [205]
Nanochannel d = 10 µm d = 10 µm 80 nm 17 ns 8 pJ [206]

Adv. Mater. 2023, 35, 2205294 2205294 (17 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

exponential dependence on ambient temperature, whereas the As illustrated in Figure 1B, nonvolatile states have been
MIT manifests a square-root defined by coupling to a dynamical chemical change such as
inducing oxygen vacancies through ionic-liquid gating with
Vcrit. ∝ TMIT − Tamb. (9) state retention spanning multiple days albeit the write/read

dynamics are slow, on the order of tens of minutes.[64,215,216]
relationship with ambient temperature.[209] By contrast, chro- Erasure is achieved through thermal annealing. Dynamical
mium-doped VO2 thin films exhibit a lower Vcrit. as compared to dopant incorporation has been used to imbue nonvolatility
undoped VO2 despite having a higher Teq..[211] Inclusion of the by annealing under H2.[162] The process takes on the order of
interstitial dopant boron in VO2 protodevices has been shown hours and is reversible upon thermal annealing. Coupling
to create a time-dependent relaxation effect that is measurable to remnant strain fields on ferroelectric substrates further
in the electrically driven MIT.[74,87] Inclusion of boron also cre- imbues some degree of nonvolatility with write/erase times on
ates pinning sites that cause the phase transition to occur in the order of minutes.[217] As will be described in a subsequent
a step-like manner over a range of applied current or voltage section, Jung et al. have demonstrated a sub-150 µs optical
values.[212] While information on the thermally driven MIT for write/erase approach that is based on accessing a metastable
doped VO2 is extensive, the electrically driven MIT of doped excited state; dynamical oscillations constitute the readout and
VO2 thin films remain relatively underexplored. can be retained for over 3 days; erasure is achieved by a cur-
rent pulse.[215] The substantial control over hysteresis widths
accessed through site-selective modification discussed in pre-
6.5. Nonvolatile Memory Storage and In-Material Processing ceding sections affords opportunities for systematically tuning
in VO2 nonvolatility through codoping with oxygen vacancies, intersti-
tial hydrogen or boron, or Li-ions.
Nonvolatile memory in neuromorphic computing is inextri- As sketched in Figure 1B, in-material processing through
cably tied to state hysteresis, and, consequently, control over dynamically reversible chemical changes induced through
transition hysteresis properties is important to neuromorphic ionic-liquid-mediated oxygen vacancy formation,[64,216] incorpo-
material design. The transition hysteresis is integral to the neu- ration of interstitial hydrogen,[162] proton- or Li-ion insertion,[65]
romorphic computing capability of VO2 because it enables on/ or photochemically induced structural transformations[215] pro-
off (R-phase/M1-phase) state retention in the absence of exciting vides a means of engendering synaptic functionality with vari-
stimuli. A narrow hysteresis requires little energetic overpo- able plasticity profiles. Control of nonvolatile memory through
tential to initiate transitions between on/off states, and, con- in-material processing serves as a means of coupling to optical
sequently, the states may be easily and efficiently overwritten. and strain fields or electrochemical modulation of volatility and
By contrast, a wider hysteresis requires greater energetic over- is a key part of the arsenal for tuning transformation character-
potential to initiate transitions, and, consequently, the mate- istics to realize the full promise of VO2 in neuromorphic archi-
rial state is retained with greater resilience against low-energy tectures.[218] Indeed, by utilizing ionic gel and electrochemical
stimulus fluctuations. VO2 is a promising candidate for integra- proton doping, multineuron VO2 transistors have achieve sev-
tion within a broad range of neuromorphic architectures owing eral synaptic functions such as pair-pulse facilitation, long-term
to its near-room-temperature MIT, but is also constrained with potentiation, long-term depression, spike-rate-dependent plas-
regards to the hysteresis width as compared to other candidates ticity, and spike-timing-dependent plasticity.[219] Another syn-
such as NbO2 characterized by a Mott transition or HfO2 where aptic transistor utilized a combination of proton intercalation
memristive behavior is underpinned by oxygen vacancy for- and oxygen vacancies to accelerate proton insertion.[66]
mation and diffusion. The challenges in imbuing nonvolatile
memory storage are a limitation to the integration of VO2 in
numerous device architectures. 6.6. VO2-Based Devices with Unconventional Device
All-thermal, nonvolatile multilevel memory has been demon- Configurations
strated as enabled by hysteretic radiative heat transfer between
deposited VO2 and its fused quartz substrate.[213] All-electrical, Distinctive device geometries and form factors can be used to
nonvolatile memory in VO2 has been enabled in practice by leverage VO2 transformation characteristics to achieve unu-
sustaining a low voltage bias, which in turn induces self-Joule sual functionality. Several strategies have been demonstrated to
heating that raises the temperature to a value sufficiently cen- reduce Vcrit. for VO2 devices. In particular, compressive strain
tered within the VO2 hysteresis.[214] At this voltage bias and con- can be used to suppress Vcrit.. A significant decrease of Vcrit.
sequent operating temperature, any M1 and R phase VO2 pre- is observed for VO2 nanobeams under compressive strain.[124]
sent is sustained. Further over- and underpotentials (≈±0.3 V) Similarly, microtube devices formed out of VO2 nanomem-
are needed to switch between low- (R) and high- (M1) resistance branes have a compressive, external strain that monotonically
states, respectively.[214] Transitions between these two states decreases the onset voltage necessary for the film to switch.[220]
can be controlled by voltage pulses (to induce the R phase) or Composite VO2–carbon nanotube devices, which utilize a
by zero-bias-voltage pulses (to return to the M1 phase). Partial carbon nanotube as a heater parallel to the direction of current
transformation within a single particle is possible by applying flow, require less energy input to initiate the phase transition.[48]
low-energy pulses deviating from the operating voltage bias, Dual-mode VO2/HfO2 neurons have been constructed and fea-
which can provide access to a range of functional resistance ture finely tunable VO2 firing thresholds (1.8–2.4 V) based on
values.[214] altering the conductance of the HfO2 dielectric layer.[221] These

Adv. Mater. 2023, 35, 2205294 2205294 (18 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

examples illustrate the potential of heterostructures with appro- and complexity of these interconnected parameters make
priately matched interfaces to selectively tune VO2 transforma- prediction of the electrical response of a VO2 device difficult.
tion characteristics. For example, some relationships between parameters are well
Conversely, several strategies have been put forth to push characterized, such as the relationship between temperature
the Vcrit. to higher values. The selective oxidation of a domain and electrical conductivity for various morphologies of VO2
has been used to elevate the transition temperature.[222] Vcrit. and doped VO2 systems. However, in a device this relationship
increases upon use of a dielectric layer between the electrical is further complicated by a feedback effect between tempera-
source and the VO2 thin film.[223] While these configura- ture, electrical conductivity of the material, power dissipated
tions may be less applicable for on-chip integration, they pro- within the device, and finally heat conducted away from the
vide valuable information on further ways to tune VO2 device active volume of the device. The time-variant modulation of
properties. these properties and their coupling can give rise to dynamical
Many intrinsic material properties play into the behavior of responses such as oscillations discussed in preceding sec-
a VO2 device, such as the electrical and thermal conductivi- tions. While certain device relationships have been character-
ties of VO2 and its distinctive set of MIT characteristics such ized, such as the relationship between Teq. and Vcrit. of a device,
as Teq., hysteresis, conductance contrast, and sharpness of other relationships are less intuitive to disentangle, such as the
onset (Figure 1). However, extrinsic device properties such as relationship between the load resistance of the circuit and the
the geometry of the device and the matrix in which the active formation of spatial inhomogeneities. To further investigate
VO2 volume is embedded play an equally important role in gov- the full capabilities of VO2 devices, the interconnected relation-
erning the response of the device to electrical stimuli (Figure 4). ships between intrinsic and extrinsic device and circuit param-
Finally, circuit components such as load resistors, extrinsic or eters must be mechanistically understood and deterministically
intrinsic capacitance of the system, type of applied electrical controlled.
bias and its time-dependence (applied voltage vs applied cur- To be technologically relevant, VO2 thin film devices must
rent; steady-state, quasi-steady, or dynamical electrical measure- be developed with increasingly small device dimensions. They
ments) also alter the device response (Figure 12). The number must require as little energy input as possible to achieve their

Figure 12. Hierarchical interactions between intrinsic material properties and phase change behavior, device parameters, and circuit level parameters.
Critical design variables exist at each level of the hierarchy, which impact the ultimate electrical response of the system. a) Tonic spiking and tonic
bursting behavior resulting from different circuit capacitance and resistance. Reproduced with permission.[8] Copyright 2018, Springer Nature. b) Crit-
ical transformation voltage as a function of device size and ambient temperature. Reproduced with permission.[196] Copyright 2014, AIP Publishing.
c) Intrinsic current–voltage response of VO2 thin films, as a result of deposition parameters, which impact film microstructure, defect concentrations
and other structural film aspects. Reproduced with permission.[204] Copyright 2013, IEEE.

Adv. Mater. 2023, 35, 2205294 2205294 (19 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

desired device performance. They will have high operating tem- crystals of VO2, suggesting fluence is a leading tunable para-
peratures and fast response times. Key areas of future devel- meter for dynamically modulating localized electronic transi-
opment include quantifying the relationships between device tions upon optical excitation.[227] Optical excitation affords a
properties and parameters and desired electrical performance, means of addressing a primary drawback of VO2 as a mem-
as well as devising a method for increasing the operational ristor or oscillator network, difficulties in imbuing nonvolatile
temperature of the device. Further investigation into the time- memory. Recent work has demonstrated that optical excitation
dependence of material properties and device behaviors charac- can be used to stabilize a putative metastable excited state that
teristics, such as switching times, time-dependent heat transfer persists at room temperature and under ambient conditions
from a device, and frequency of device switching, is essential. even without continued photoexcitation or electrical bias. The
As both intrinsic and extrinsic properties are at play, an exhaus- optical memory is addressed and read at <150 µs times in the
tive characterization of these relationships requires both experi- form of voltage oscillations between the insulating and metallic
mental and computational approaches bridging across decades states.[215]
of length and time scales. All-optical infrared-switching VO2 devices in relatively simple
configurations feature coupling to Si waveguides.[229,230] Under
voltage bias, VO2/Si devices can operate under low energy
7. Optically Driven MIT, Associated Structural requirements (as low as 23.5 pJ per pulse) for writing memory
with good retention (over 200 h) of the memory state.[230] Repro-
Rearrangements in VO2, and Implications for
ducibility for such devices has been shown over 10 000 write/
Neuromorphic Perception Systems erase instances.[230] Further VO2 optical device adaptations have
been designed using metasurfaces to control optical coupling
The burgeoning role of photonic memristors in neuromor- through the use of bespoke plasmonic meta-atoms.[231,232] Au
phic computing architectures is an exciting prospect recently was shown an advantageous metaelement to align surface
reviewed by Goi et al.[224] The optically triggered MIT of spe- plasmon resonance wavelengths to photoexcitation sources in
cifically VO2 is notable for its rapid (as fast as ≈26 fs), single- a study utilizing a perforated top film that achieved 50% power
photon Mott–Hubbard transition from an insulating state to reduction to induce the VO2 MIT.[231] A related device advanced
a metallic state.[225] Ultrafast electron diffraction coupled to by integration to a Ti/Au/VO2/Ti/Au threshold switching
time-resolved infrared transmittance measurements reveals selector enabled promising control of MIT transition charac-
the emergence of a metallic state with a distinctive dielectric teristics and ultimately produced an optical system capable of
signature (optical bandgap below 0.25 eV) while a metastable image processing tasks.[232] Material pairings with VO2 to form
M1 structure of VO2 persists.[226] Radial pair distribution differ- p–n heterojunctions have been realized in WSe2/VO2 hetero-
ence function measurements indicate valence charge density structure devices capable of switching between photovoltaic
increases in the V–V dimerized bond (derived from dxy sub- and photoconductive detection modes upon transformation
shell interactions) upon photoexcitation, concomitant with dxz from insulating to metallic phases, respectively.[233]
subshell (associated with octahedrally chained V and O atoms) A mixed optical/CMOS-integrated device has been designed
charge density reduction.[226] This indicates an electronic tran- based on three VO2 oscillators to construct an optical rec-
sition in VO2 occurs during the optical MIT prior to the full ognition system.[234] The circuit stores training data as cou-
structural rearrangement from M1 to R phase and furthermore pling weights of the VO2 oscillators. The image sought to be
suggests that the latent heat of the thermally activated MIT is examined is encoded in the form of time delays in oscillator
largely necessary for the structural transition rather than elec- inputs; phase differentials in oscillator waveforms are used to
tronic transition.[226] reconstruct the image. A particularly compelling demonstra-
Photoexcitation fluence provides a convenient control over tion of the utility of optically coupled VO2 devices comes from
induction of either solely electronic versus a coupled electronic the work of Yang and co-workers who recently demonstrated a
and structural transition.[226,227] Morrison et al. found that at spike-based neuromorphic perception module based on high-
800 nm wavelength excitation, a pump fluence >9 mJ cm−2 quality epitaxial VO2 thin films with high conductance con-
was sufficient to trigger structural reordering; electronic tran- trast and sharp onset temperatures. By coupling the nonlinear
sitions to a more metallized configuration were accessible at VO2 device element with calibration and scaling resistors and
pump fluences as low as 2 mJ cm−2.[226] At the higher fluence, an arrays of sensors, the researchers were able to encode illu-
disrupting the atomic ordering to yield R phase occurred over minance, temperature, pressure, and curvature signals in the
≈300 ps.[226] Once formed, the metallic VO2 state is stabilized spiking response, enabling identification of hand gestures as
by ligand-to-metal charge-transfer-type transitions (O 2p to V 3d required for haptics applications with the use of spiking neural
bands), allowing the state to persist on a comparatively long- networks.[235]
lived picosecond timescale before relaxing back to the insu-
lating electronic state (and M1 phase).[225] The longevity of the
optically transitioned state is posited to electron–phonon carrier 8. Conclusions and Outlook
relaxation pathways.[226] The pump fluence furthermore influ-
ences domain heterogeneity during the photoinduced transi- A four-to-five orders of magnitude improvement in energy con-
tion, yielding insulating and metallic domains spatially sepa- sumed and time-to-solution for the same problem has been
rated on scales of ≈100 nm.[225,228] Similar fluence-dependent projected for neuromorphic computing over current digital
inhomogeneity was observed even within individual single computing based on the design of new circuit elements that

Adv. Mater. 2023, 35, 2205294 2205294 (20 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

leverage Mott transitions, embedding a nontrivial number of measurements across the field is imperative to benchmark
such circuit elements in CMOS-integrated networks, and oper- accessible limits of energy efficiency and speed-up.
ating them using algorithms purpose-designed for nondigital From the perspective of inverse design, several questions
devices.[15,236,237] The predicted improvements result from remain, for instance, 1) a priori prediction of the temporal
computations being performed in an entirely different way characteristics of VO2 circuit elements, e.g., periodic versus
from a digital platform at all levels of the stack. Beyond energy bursting versus chaotic,[239] or 2) additional dynamical degrees
efficiency, considerable excitement about the promise of neu- of freedom that can be enabled within a network integrating a
romorphic computing arises from prospects of realizing alto- nontrivial ensemble of circuit elements as opposed to an indi-
gether new function. Brain-inspired computing holds promise vidual device. A growing body of experimental work shows that
for integrating sensing capabilities with logic and memory, these factors are closely interrelated.[236] An attractive approach
thereby enabling 3D spatiotemporal perception in a manner to address the complexity of inverse design of neuromorphic
available to biological organisms. This could take the form of device elements across scales is to leverage carefully con-
relatively simple edge computing to determine the most valu- structed libraries with site-selective modification and systematic
able information for further retention and processing or imple- variation of circuit parameters, and with the help of Bayesian
mentation in autonomous vehicles, robotic systems, or the idea machine learning (ML) and multimodal operando measure-
of “complementary intelligence” bridging human and machine ments, connect target neuronal behavior to the necessary mate-
intelligence.[238] rial properties. To navigate the vast available design space, it
Realizing the full promise of neuromorphic computing will be important to design algorithms to express the physical/
requires deciphering and precisely modulating the mechanisms chemical features of materials that best underpin a specific
of electronic instabilities that are at the heart of low-entropy neuronal function. Such an approach can address challenges
electronic transitions such as the Mott–Peierls transition of with inverse design resulting from the large dimensionality of
VO2. Approaching the fundamental limits of energy efficiency the feature space, the sparsity of available data, and the fact that
and speed possible using VO2 as an active neuromorphic ele- the region of interest with desirable properties is often a narrow
ment necessitates control over the specific compositional, struc- singularity. For instance, such an approach can inform the
tural, and mesoscale levers for tuning phase stabilities, trans- design of codoping schemes (Equations (4)–(6)) or guide the
formation pathways, and nucleation mechanisms illustrated use of a specific directional strain field to couple with a specific
in Figure 4. In this article, we have reviewed current under- doped VO2 film under certain geometric constraints (Figure 7).
standing of the role of nonlinear dynamical transitions in neu- Strain plays a key role in altering the electronic properties of
romorphic computing, tracing the connections between neu- VO2. First, strain couples directly to electronic transformations
ronal and synaptic function and intrinsic material properties by modifying VO2’s lattice and electronic structure. Strain is also
and their coupling through interfaces with photoexcitation and ever-present in realistic device structures, e.g., in residual strains
voltage, temperature, and strain fields (Figure 2). We have fur- induced during fabrication, such as those stemming from epi-
ther examined current ideas of transformation mechanisms in taxy. As such, electronic properties of VO2 can potentially be tai-
VO2 and opportunities for modifying the transformation char- lored by displacing atoms through controlled mechanical strain,
acteristics of this material by site-selective modification (sub- either external strain applied dynamically, e.g., via a piezoelectric
stitutional or interstitial doping), defects, and strain (external, actuator, or built-in statically, e.g., via epitaxy. Particularly tan-
chemical, or epitaxial). A curated set of examples from the lit- talizing, these strains may directly induce a phase transforma-
erature have been used to describe processing–structure–prop- tion (e.g., by piezoelectric actuation) or may poise VO2 at the
erty relationships for single-crystalline nanostructures and VO2 cusp of a phase transformation or electrical instability, thereby
films with a specific emphasis on the role of boundary condi- dramatically reducing the energy and time required to switch
tions. In addition to the modification of MIT characteristics by states. However, fully harnessing these effects in a deterministic
substrate coupling, we have highlighted the dynamical evolu- manner requires a deeper mechanistic understanding of the
tion of stress across the MIT in the context of device resilience. mechanoelectronic interplay in VO2 and is thus a key topic of
Figure 13 illustrates open questions and research oppor- future studies. For instance, does strain induced by doping of
tunities related to understanding and modulating the MIT of VO2 directly influence electronic switching (e.g., from the strain
VO2 across decades of length and time scales through con- level itself) or are these effects more indirect, e.g., by the dopant
trol of material properties, as well as modes and strengths of serving as a nucleation site for the transformation?
coupling to external composition, strain, thermal, and electric Another emerging frontier relates to the design of complex
fields (Figure 2). A key gap, outlined in Section 2, pertains to ternary or quaternary vanadium oxides with precisely tunable
the limited current knowledge of how to map specific neuro- charge- and spin-ordering motifs that afford a still broader
morphic function to intrinsic material, compositional, and range of phase transformation mechanisms (Figure 13, right).
interface properties. As noted above, channel formation means In MxV2O5 and MxM′yV2O5, where M and M′ represent s-, p-,
transport in devices is nonuniform, and first-order transitions or d-block ions, x and y are their respective stoichiometries,
are accompanied by some extent of hysteresis. A critical need and V2O5 represents one of several possible polymorphs, the
is to derive atomistic understanding of the dynamical evolution manifestation of mechanisms such as polaron oscillation, lone-
of atomic connectivity and electronic structure in active chan- pair anharmonicity, and canonical Mott transitions provide a
nels using methods such as operando scanning transmission basis for conductance switching across a broad range of crit-
electron microscopy and X-ray ptychography in conjunction ical temperature, current, and voltage values.[7,240,241] In these
with electrothermal measurements. Standardization of such compounds, Li-ion or proton intercalation can bring about

Adv. Mater. 2023, 35, 2205294 2205294 (21 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

Figure 13. Challenges and opportunities for designing neuromorphic elements from VO2. The microscopic response of VO2 to an external stimulus
(electric field, temperature gradient, elastic strain, etc.) results from a complex interaction between intrinsic material properties, properties of sur-
rounding materials (e.g., contacts and substrate) and interfaces, and the external stimulus itself. Predicting the response of a material system (and
designing a material to exhibit a particular response) across multiple decades of length and timescales requires multiscale, multiphysics models
informed by operando measurements of spatiotemporally resolved electronic structure and electrothermal transport. External stimuli operate on a
material to change its internal structure, which in turn alters the microscopic properties of a particular volume of material. These changes in structure
occur across multiple length-scales. Chemistry-aware machine learning techniques can be used in conjunction with analytical frameworks and oper-
ando measurements to accurately predict a material response to a particular set of external fields. Lower-half, top, left panel reproduced with permis-
sion.[71] Copyright 2017, Springer Nature; lower-half, top, middle panel reproduced with permission.[40] Copyright 2018, American Chemical Society;
lower-half, bottom, left and middle panels reproduced with permission.[243] Copyright 2020, Elsevier.

reconfigurability across different regions of the phase diagram, Acknowledgements

enabling toggling between neuronal (thresholding, amplifica-
This research was primarily supported by Texas A&M X-Grants. S.B,
tion, integration) and synaptic (learning, memory) modes.[242] P.J.S., P.S., A.B., and R.M.G. acknowledge the Data-Enabled Design and
Discovery of Energy Materials fellowship funded under National Science
Foundation (Award No. DGE-1545403). This material is based upon
work supported by the National Science Foundation under Grant No.
Supporting Information DMR-1847956 (P.J.S.). S.B. further acknowledges support from the Welch
Supporting Information is available from the Wiley Online Library or Foundation. The authors thank David Santos and Joseph V. Handy for
from the author. assistance in figure preparation.

Adv. Mater. 2023, 35, 2205294 2205294 (22 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

Conflict of Interest [26] S. D. Ha, S. Ramanathan, J. Appl. Phys. 2011, 110, 071101.
[27] A. M. Ionescu, H. Riel, Nature 2011, 479, 329.
The authors declare no conflict of interest. [28] Z. Yang, C. Ko, S. Ramanathan, Annu. Rev. Mater. Res. 2011, 41,
337.
[29] S. Kumar, X. Wang, J. P. Strachan, Y. Yang, W. D. Lu, Nat. Rev.
Mater. 2022, 7, 575.
Keywords [30] M. D. Pickett, G. Medeiros-Ribeiro, R. S. Williams, Nat. Mater.
doping, electronic structures, metal–insulator transition, Mott transition, 2013, 12, 114.
neuromorphic computing, phase transition, vanadium dioxide [31] M. D. Pickett, R. S. Williams, Nanotechnology 2012, 23, 215202.
[32] Z. Wang, M. Rao, J. W. Han, J. Zhang, P. Lin, Y. Li, C. Li, W. Song,
Received: June 12, 2022 S. Asapu, R. Midya, Y. Zhuo, H. Jiang, J. H. Yoon, N. K. Upadhyay,
Revised: August 2, 2022 S. Joshi, M. Hu, J. P. Strachan, M. Barnell, Q. Wu, H. Wu, Q. Qiu,
Published online: November 29, 2022 R. S. Williams, Q. Xia, J. J. Yang, Nat. Commun. 2018, 9, 3208.
[33] S. Kumar, J. P. Strachan, R. S. Williams, Nature 2017, 548, 318.
[34] T. J. Park, S. Deng, S. Manna, A. N. M. N. Islam, H. Yu,
Y. Yuan, D. D. Fong, A. A. Chubykin, A. Sengupta,
[1] R. Vinuesa, H. Azizpour, I. Leite, M. Balaam, V. Dignum, S. K. R. S. Sankaranarayanan, S. Ramanathan, Adv. Mater. 2022,
S. Domisch, A. Felländer, S. D. Langhans, M. Tegmark, F. F. Nerini, 2203352.
Nat. Commun. 2020, 11, 233. [35] S. Wall, S. Yang, L. Vidas, M. Chollet, J. M. Glownia, M. Kozina,
[2] N. C. Thompson, K. Greenewald, K. Lee, G. F. Manso, IEEE Spec- T. Katayama, T. Henighan, M. Jiang, T. A. Miller, D. A. Reis,
trumSeptember 2021. L. A. Boatner, O. Delaire, M. Trigo, Science 2018, 362, 572.
[3] R. Roscher, B. Bohn, M. F. Duarte, J. Garcke, IEEE Access 2020, 8, [36] W. H. Brito, M. C. O. Aguiar, K. Haule, G. Kotliar, Phys. Rev. Lett.
42200. 2016, 117, 056402.
[4] Cisco, Global Cloud Index: Forecast and Methodology, 2018. [37] J. Del Valle, P. Salev, S. Gariglio, Y. Kalcheim, I. K. Schuller,
[5] E. Masanet, A. Shehabi, N. Lei, S. Smith, J. Koomey, Science 2020, J. M. Triscone, Nano Lett. 2022, 22, 1251.
367, 984. [38] L. Whittaker, C. J. Patridge, S. Banerjee, J. Phys. Chem. Lett. 2011,
[6] I. K. Schuller, R. Stevens, R. Pino, M. Pechan Neuromorphic Com- 2, 745.
puting: From Materials to Systems Architecture Roundtable, USDOE [39] W. Fan, J. Cao, J. Seidel, Y. Gu, J. W. Yim, C. Barrett, K. M. Yu, J. Ji,
Office of Science, Gaithersburg, MD 2015. R. Ramesh, L. Q. Chen, J. Wu, Phys. Rev. B: Condens. Matter Mater.
[7] J. L. Andrews, D. A. Santos, M. Meyyappan, R. S. Williams, Phys. 2011, 83, 235102.
S. Banerjee, Trends Chem. 2019, 1, 711. [40] E. J. Braham, D. Sellers, E. Emmons, R. Villarreal, H. Asayesh-
[8] W. Yi, K. K. Tsang, S. K. Lam, X. Bai, J. A. Crowell, E. A. Flores, Nat. Ardakani, N. A. Fleer, K. E. Farley, R. Shahbazian-Yassar,
Commun. 2018, 9, 4661. R. Arròyave, P. J. Shamberger, S. Banerjee, Chem. Mater. 2018, 30,
[9] J. Torrejon, M. Riou, F. A. Araujo, S. Tsunegi, G. Khalsa, 214.
D. Querlioz, P. Bortolotti, V. Cros, K. Yakushiji, A. Fukushima, [41] Y. G. Liang, S. Lee, H. S. Yu, H. R. Zhang, Y. J. Liang, P. Y. Zavalij,
H. Kubota, S. Yuasa, M. D. Stiles, J. Grollier, Nature 2017, 547, 428. X. Chen, R. D. James, L. A. Bendersky, A. V. Davydov, X. H. Zhang,
[10] M. Lundstrom, Science 2003, 299, 210. I. Takeuchi, Nat. Commun. 2020, 11, 3539.
[11] G. M. Moore, Electronics 1965, 38, 114. [42] J. D. Budai, J. Hong, M. E. Manley, E. D. Specht, C. W. Li,
[12] R. H. Dennard, F. H. Gaensslen, H.-N. Yu, V. L. Rideout, J. Z. Tischler, D. L. Abernathy, A. H. Said, B. M. Leu, L. A. Boatner,
E. Bassous, A. R. LeBlanc, IEEE J. Solid-State Circuits 1974, 9, 256. R. J. McQueeney, O. Delaire, Nature 2014, 515, 535.
[13] M. B. Taylor, IEEE Micro. 2013, 33, 8. [43] Y. Cui, Y. Ke, C. Liu, Z. Chen, N. Wang, L. Zhang, Y. Zhou, S. Wang,
[14] J. Backus, Commun. ACM 1978, 21, 613. Y. Gao, Y. Long, Joule 2018, 2, 1707.
[15] J. D. Kendall, S. Kumar, Appl. Phys. Rev. 2020, 7, 029901. [44] X. Cao, T. Chang, Z. Shao, F. Xu, H. Luo, P. Jin, Matter 2020, 2,
[16] A. Calimera, E. Macii, M. Poncino, Funct. Neurol. 2013, 28, 191. 862.
[17] S. K. Esser, P. A. Merolla, J. V. Arthur, A. S. Cassidy, R. Appuswamy, [45] S. Lany, J. Phys.: Condens. Matter 2015, 27, 283203.
A. Andreopoulos, D. J. Berg, J. L. McKinstry, T. Melano, [46] S. Choi, A. Kutepov, K. Haule, M. van Schilfgaarde, G. Kotliar, Nat.
D. R. Barch, C. Di Nolfo, P. Datta, A. Amir, B. Taba, M. D. Flickner, Commun. 2016, 1, 16001.
D. S. Modha, Proc. Natl. Acad. Sci. USA 2016, 113, 11441. [47] X. Yuan, Y. Zhang, T. A. Abtew, P. Zhang, W. Zhang, Phys. Rev. B:
[18] G. Csaba, W. Porod, Appl. Phys. Rev. 2020, 7, 011302. Condens. Matter Mater. Phys. 2012, 86, 235103.
[19] S. J. Kim, S. B. Kim, H. W. Jang, iScience 2021, 24, 101889. [48] S. M. Bohaichuk, S. Kumar, G. Pitner, C. J. McClellan, J. Jeong,
[20] K. Amunts, C. Ebell, J. Muller, M. Telefont, A. Knoll, T. Lippert, M. G. Samant, H. S. P. Wong, S. S. P. Parkin, R. S. Williams,
Neuron 2016, 92, 574. E. Pop, Nano Lett. 2019, 19, 6751.
[21] E. J. Fuller, F. El Gabaly, F. Léonard, S. Agarwal, S. J. Plimpton, [49] S. Kumar, R. S. Williams, Nat. Commun. 2018, 9, 2030.
R. B. Jacobs-Gedrim, C. D. James, M. J. Marinella, A. A. Talin, Adv. [50] B. K. Ridley, Proc. Phys. Soc. 1963, 82, 954.
Mater. 2017, 29, 1604310. [51] T. D. Brown, S. Kumar, R. S. Williams, Appl. Phys. Rev. 2022, 9,
[22] C. Sen Yang, D. S. Shang, N. Liu, E. J. Fuller, S. Agrawal, A. A. Talin, 011308.
Y. Q. Li, B. G. Shen, Y. Sun, Adv. Funct. Mater. 2018, 28, 1804170. [52] A. Ascoli, A. S. Demirkol, R. Tetzlaff, S. Slesazeck, T. Mikolajick,
[23] P. A. Merolla, J. V. Arthur, R. Alvarez-icaza, A. S. Cassidy, J. Sawada, L. O. Chua, Front. Neurosci. 2021, 15, 651452.
F. Akopyan, B. L. Jackson, N. Imam, C. Guo, Y. Nakamura, [53] A. Ascoli, S. Slesazeck, H. Mähne, R. Tetzlaff, S. Member,
B. Brezzo, I. Vo, S. K. Esser, R. Appuswamy, B. Taba, A. Amir, T. Mikolajick, S. Member, IEEE Trans. Circuits Syst. I: Regul. Pap.
M. D. Flickner, W. P. Risk, R. Manohar, D. S. Modha, Science 2014, 2015, 62, 1165.
345, 668. [54] I. Messaris, T. D. Brown, A. S. Demirkol, A. Ascoli, M. M. Al Chawa,
[24] N. Savage, Commun. ACM 2017, 60, 13. R. S. Williams, S. Member, R. Tetzlaff, S. Member, L. O. Chua,
[25] R. W. Keyes, Rep. Prog. Phys. 2005, 68, 2701. IEEE Trans. Circuits Syst. I Regul. Pap. 2021, 68, 4979.

Adv. Mater. 2023, 35, 2205294 2205294 (23 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

[55] L. Chua, IEEE Trans. Circuit Theory 1971, CT-18, 507. [86] K. Post, R. G. Robins, Electrochim. Acta 1976, 21, 401.
[56] L. O. Chua, S. M. Kang, Proc. IEEE 1976, 64, 209. [87] T. E. G. Alivio, D. G. Sellers, H. Asayesh-Ardakani, E. J. Braham,
[57] G. A. Gibson, Adv. Funct. Mater. 2018, 28, 1704175. G. A. Horrocks, K. E. Pelcher, R. Villareal, L. Zuin, P. J. Shamberger,
[58] A. S. Alexandrov, A. M. Bratkovsky, B. Bridle, S. E. Savel’Ev, R. Arroyave, R. Shahbazian-Yassar, S. Banerjee, Chem. Mater. 2017,
D. B. Strukov, R. Stanley Williams, Appl. Phys. Lett. 2011, 99, 29, 5401.
202104. [88] C. J. Patridge, L. Whittaker, B. Ravel, S. Banerjee, J. Phys. Chem. C
[59] L. O. Chua, IEEE Trans. Circuits Syst. I: Fundam. Theory Appl. 1999, 2012, 116, 3728.
46, 71. [89] J. Du, Y. Gao, H. Luo, L. Kang, Z. Zhang, Z. Chen, C. Cao, Sol.
[60] A. Yano, H. Clarke, D. G. Sellers, E. J. Braham, T. E. G. Alivio, Energy Mater. Sol. Cells 2011, 95, 469.
S. Banerjee, P. J. Shamberger, J. Phys. Chem. C 2020, 124, 21223. [90] H. Futaki, M. Aoki, Hisao Futaki, Minoru Aoki, Jpn. J. Appl. Phys.
[61] P. Schofield, E. J. Braham, B. Zhang, J. L. Andrews, H. K. Drozdick, 1969, 8, 1008.
D. Zhao, W. Zaheer, R. M. Gurrola, K. Xie, P. J. Shamberger, [91] L. Whittaker, T.-L. Wu, C. J. Patridge, G. Sambandamurthy,
X. Qian, S. Banerjee, Chem. Commun. 2022, 58, 6586. S. Banerjee, J. Mater. Chem. 2011, 21, 5580.
[62] D. Wegkamp, J. Stähler, Prog. Surf. Sci. 2015, 90, 464. [92] S. J. Liu, H. W. Fang, Y. T. Su, J. H. Hsieh, Jpn. J. Appl. Phys. 2014,
[63] D. Lee, B. Chung, Y. Shi, G. Kim, N. Campbell, F. Xue, K. Song, 53, 063201.
S. Choi, J. P. Podkaminer, T. H. Kim, P. J. Ryan, J. Kim, T. R. Paudel, [93] W. Lu, G. Zhao, B. Song, J. Li, X. Zhang, G. Han, Surf. Coat.
J. Kang, J. W. Spinuzzi, D. A. Tenne, E. Y. Tsymbal, M. S. Rzchowski, Technol. 2017, 320, 311.
L. Q. Chen, J. Lee, C. B. Eom, Science 2018, 362, 1037. [94] C. Piccirillo, R. Binions, I. P. Parkin, Eur. J. Inorg. Chem. 2007, 2007,
[64] J. Jeong, N. Aetukuri, T. Graf, T. D. Schladt, M. G. Samant, 4050.
S. S. P. Parkin, Science 2013, 339, 1402. [95] A. Krammer, O. Bouvard, A. Schüler, Energy Procedia 2017, 122,
[65] J. Park, C. Oh, J. Son, J. Mater. Chem. C 2021, 9, 2521. 745.
[66] C. Oh, I. Kim, J. Park, Y. Park, M. Choi, J. Son, Adv. Electron. Mater. [96] Y. Gao, C. Cao, L. Dai, H. Luo, M. Kanehira, Y. Ding, Z. L. Wang,
2021, 7, 2000802. Energy Environ. Sci. 2012, 5, 8708.
[67] Z. Zhang, R. D. James, S. Müller, Acta Mater. 2009, 57, 4332. [97] W. Burkhardt, T. Christmann, B. K. Meyer, W. Niessner, D. Schalch,
[68] A. R. Balakrishna, J. Electrochem. Energy Convers. Storage 2022, 19, A. Scharmann, Thin Solid Films 1999, 345, 229.
040802. [98] L. Dai, S. Chen, J. Liu, Y. Gao, J. Zhou, Z. Chen, C. Cao, H. Luo,
[69] S. A. Howard, E. Evlyukhin, G. Páez Fajardo, H. Paik, D. G. Schlom, M. Kanehira, Phys. Chem. Chem. Phys. 2013, 15, 11723.
L. F. J. Piper, Adv. Mater. Interfaces 2021, 8, 2001790. [99] I. Top, J. Schläfer, R. Binions, I. Papakonstantinou,
[70] N. F. Quackenbush, J. W. Tashman, J. A. Mundy, S. Sallis, H. Paik, S. Srimurugananthan, M. Powell, I. Parkin, C. J. Carmalt,
R. Misra, J. A. Moyer, J. H. Guo, D. A. Fischer, J. C. Woicik, I. Abrahams, MRS Adv. 2018, 3, 1863.
D. A. Muller, D. G. Schlom, L. F. J. Piper, Nano Lett. 2013, 13, 4857. [100] Y. Cui, Y. Wang, B. Liu, H. Luo, Y. Gao, RSC Adv. 2016, 6, 64394.
[71] B. M. Hudak, S. W. Depner, G. R. Waetzig, A. Talapatra, [101] M. Wan, M. Xiong, N. Li, B. Liu, S. Wang, W. Y. Ching, X. Zhao,
R. Arroyave, S. Banerjee, B. S. Guiton, Nat. Commun. 2017, 8, Appl. Surf. Sci. 2017, 410, 363.
15316. [102] K. Kawatani, T. Kanki, H. Tanaka, Phys. Rev. B: Condens. Matter
[72] A. Parija, G. R. Waetzig, J. L. Andrews, S. Banerjee, J. Phys. Chem. Mater. Phys. 2014, 90, 054203.
C 2018, 122, 25709. [103] P. Baum, D. S. Yang, A. H. Zewail, Springer Ser. Chem. Phys. 2009,
[73] A. Krammer, A. Magrez, W. A. Vitale, P. Mocny, P. Jeanneret, 92, 116.
E. Guibert, H. J. Whitlow, A. M. Ionescu, A. Schüler, J. Appl. Phys. [104] H. Clarke, B. D. Caraway, D. G. Sellers, E. J. Braham, S. Banerjee,
2017, 122, 045304. R. Arróyave, P. J. Shamberger, Phys. Rev. Mater. 2018, 2, 103402.
[74] D. G. Sellers, E. J. Braham, R. Villarreal, B. Zhang, A. Parija, [105] J. Zhou, Y. Gao, X. Liu, Z. Chen, L. Dai, C. Cao, H. Luo,
T. D. Brown, T. E. G. Alivio, H. Clarke, L. R. De Jesus, L. Zuin, M. Kanahira, C. Sun, L. Yan, Phys. Chem. Chem. Phys. 2013, 15,
D. Prendergast, X. Qian, R. Arroyave, P. J. Shamberger, S. Banerjee, 7505.
J. Am. Chem. Soc. 2020, 142, 15513. [106] F. Béteille, R. Morineau, J. Livage, M. Nagano, Mater. Res. Bull.
[75] C. E. Wilson, A. E. Gibson, J. J. Argo, P. A. Loughney, W. Xu, 1997, 32, 1109.
G. King, V. Doan-Nguyen, J. Mater. Res. 2021, 36, 268. [107] S. Chen, J. Liu, L. Wang, H. Luo, Y. Gao, J. Phys. Chem. C 2014, 118,
[76] E. J. Braham, J. L. Andrews, T. E. G. Alivio, N. A. Fleer, S. Banerjee, 18938.
Phys. Status Solidi A 2018, 215, 1700884. [108] E. Cavanna, J. P. Segaud, J. Livage, Mater. Res. Bull. 1999, 34, 167.
[77] M. Marezio, D. B. McWhan, J. P. Remeika, P. D. Dernier, Phys. Rev. [109] Y. Wu, L. Fan, Q. Liu, S. Chen, W. Huang, F. Chen, G. Liao, C. Zou,
B 1972, 5, 2541. Z. Wu, Sci. Rep. 2015, 5, 9328.
[78] A. Zylbersztejn, N. F. Mott, Phys. Rev. B 1975, 11, 4383. [110] R. Lopez, L. A. Boatner, T. E. Haynes, L. C. Feldman, R. F. Haglund,
[79] A. Tselev, I. A. Luk’Yanchuk, I. N. Ivanov, J. D. Budai, J. Z. Tischler, J. Appl. Phys. 2002, 92, 4031.
E. Strelcov, A. Kolmakov, S. V. Kalinin, Nano Lett. 2010, 10, 4409. [111] L. Whittaker, C. Jaye, Z. Fu, D. A. Fischer, S. Banerjee, J. Am. Chem.
[80] J. L. Victor, M. Gaudon, G. Salvatori, O. Toulemonde, N. Penin, Soc. 2009, 131, 8884.
A. Rougier, J. Phys. Chem. Lett. 2021, 12, 7792. [112] H. Asayesh-Ardakani, W. Yao, A. Nie, P. M. Marley, E. Braham,
[81] Y. Filinchuk, N. A. Tumanov, V. Ban, H. Ji, J. Wei, M. W. Swift, R. F. Klie, S. Banerjee, R. Shahbazian-Yassar, Appl. Phys. Lett. 2017,
A. H. Nevidomskyy, D. Natelson, J. Am. Chem. Soc. 2014, 136, 109, 233104.
8100. [113] K. Miyazaki, K. Shibuya, M. Suzuki, H. Wado, A. Sawa, Jpn. J. Appl.
[82] C. H. Griffiths, H. K. Eastwood, J. Appl. Phys. 1974, 45, 2201. Phys. 2014, 53, 071102.
[83] D. Ruzmetov, S. D. Senanayake, V. Narayanamurti, S. Ramanathan, [114] M. K. Dietrich, F. Kuhl, A. Polity, P. J. Klar, Appl. Phys. Lett. 2017,
Phys. Rev. B 2008, 77, 195442. 110, 141907.
[84] J. Montero, Y.-X. Ji, S.-Y. Li, G. A. Niklasson, C. G. Granqvist, J. Vac. [115] K. Miyazaki, K. Shibuya, M. Suzuki, K. Sakai, J.-I. Fujita, A. Sawa,
Sci. Technol., B: Nanotechnol. Microelectron.: Mater. Process., Meas., AIP Adv. 2016, 6, 55012.
Phenom. 2015, 33, 031805. [116] C. Ji, Z. Wu, L. Lu, X. Wu, J. Wang, X. Liu, H. Zhou, Z. Huang,
[85] G. A. Horrocks, S. Singh, M. F. Likely, G. Sambandamurthy, J. Gou, Y. Jiang, J. Mater. Chem. C 2018, 6, 6502.
S. Banerjee, ACS Appl. Mater. Interfaces 2014, 6, 15726. [117] Y. Shi, L.-Q. Chen, ArXiv 2020, arXiv:2001.02264v1.

Adv. Mater. 2023, 35, 2205294 2205294 (24 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

[118] Y. Shi, F. Xue, L.-Q. Chen, EPL 2017, 120, 46003. [146] J. Li, J. Dho, J. Cryst. Growth 2010, 312, 3287.
[119] N. I. Cool, D. G. Sellers, M. Al-Hashimi, S. Banerjee, ACS Appl. [147] E. Breckenfeld, H. Kim, K. Burgess, N. Charipar, S. F. Cheng,
Energy Mater. 2022, 5, 4829. R. Stroud, A. Pique, ACS Appl. Mater. Interfaces 2017, 9, 1577.
[120] J. Nag, R. F. Haglund, J. Phys.: Condens. Matter 2008, 20, 264016. [148] Y. Muraoka, K. Saeki, R. Eguchi, T. Wakita, M. Hirai, T. Yokoya,
[121] K. Nagashima, T. Yanagida, H. Tanaka, T. Kawai, J. Appl. Phys. S. Shin, J. Appl. Phys. 2011, 109, 043702.
2007, 101, 026103. [149] M. Liu, M. Wagner, J. Zhang, A. McLeod, S. Kittiwatanakul, Z. Fei,
[122] R. E. Marvel, R. R. Harl, V. Craciun, B. R. Rogers, R. F. Haglund, E. Abreu, M. Goldflam, A. J. Sternbach, S. Dai, K. G. West, J. Lu,
Acta Mater. 2015, 91, 217. S. A. Wolf, R. D. Averitt, D. N. Basov, Appl. Phys. Lett. 2014, 104,
[123] J. Narayan, V. M. Bhosle, J. Appl. Phys. 2006, 100, 103524. 121905.
[124] J. Cao, E. Ertekin, V. Srinivasan, W. Fan, S. Huang, H. Zheng, [150] M. Liu, A. J. Sternbach, M. Wagner, T. V. Slusar, T. Kong,
J. W. L. Yim, D. R. Khanal, D. F. Ogletree, J. C. Grossman, J. Wu, S. L. Bud’ko, S. Kittiwatanakul, M. M. Qazilbash, A. McLeod,
Nat. Nanotechnol. 2009, 4, 732. Z. Fei, E. Abreu, J. Zhang, M. Goldflam, S. Dai, G. X. Ni, J. Lu,
[125] J. Wu, Q. Gu, B. S. Guiton, N. P. De Leon, L. Ouyang, H. Park, H. A. Bechtel, M. C. Martin, M. B. Raschke, R. D. Averitt,
Nano Lett. 2006, 6, 2313. S. A. Wolf, H. T. Kim, P. C. Canfield, D. N. Basov, Phys. Rev. B: Con-
[126] J. Yoon, H. Kim, X. Chen, N. Tamura, B. S. Mun, C. Park, H. Ju, dens. Matter Mater. Phys. 2015, 91, 245155.
ACS Appl. Mater. Interfaces 2016, 8, 2280. [151] M. Liu, A. J. Sternbach, D. N. Basov, Rep. Prog. Phys. 2017, 80,
[127] A. C. Jones, S. Berweger, J. Wei, D. Cobden, M. B. Raschke, Nano 014501.
Lett. 2010, 10, 1574. [152] E. Abreu, M. Liu, J. Lu, K. G. West, S. Kittiwatanakul, W. Yin,
[128] A. Frenzel, M. M. Qazilbash, M. Brehm, B.-G. Chae, B.-J. Kim, S. A. Wolf, R. D. Averitt, New J. Phys. 2012, 14, 083026.
H.-T. Kim, A. V. Balatsky, F. Keilmann, D. N. Basov, Phys. Rev. B: [153] C. B. Greenberg, Thin Solid Films 1983, 110, 73.
Condens. Matter Mater. Phys. 2009, 80, 115115. [154] P. Jin, S. Tanemura, Thin Solid Films 1996, 281–282, 239.
[129] J. H. Park, J. M. Coy, T. Serkan Kasirga, C. Huang, Z. Fei, S. Hunter, [155] P. Jin, S. Tanemura, Jpn. J. Appl. Phys. 1995, 34, 2459.
D. H. Cobden, Nature 2013, 500, 431. [156] P. Jin, S. Nakao, S. Tanemura, Thin Solid Films 1998, 324, 151.
[130] B. Lazarovits, K. Kim, K. Haule, G. Kotliar, Phys. Rev. B: Condens. [157] Z. P. Wu, A. Miyashita, S. Yamamoto, H. Abe, I. Nashiyama,
Matter Mater. Phys. 2010, 81, 115117. K. Narumi, H. Naramoto, J. Appl. Phys. 1999, 86, 5311.
[131] Y. Muraoka, Z. Hiroi, Appl. Phys. Lett. 2002, 80, 583. [158] J. C. Lee, G. V. Jorgenson, Sol. Energy Mater. 1986, 14, 205.
[132] N. F. Quackenbush, H. Paik, M. J. Wahila, S. Sallis, M. E. Holtz, [159] P. Jin, S. Nakao, S. Tanemura, Nucl. Instrum. Methods Phys. Res.,
X. Huang, A. Ganose, B. J. Morgan, D. O. Scanlon, Y. Gu, F. Xue, Sect. B 1998, 141, 419.
L. Q. Chen, G. E. Sterbinsky, C. Schlueter, T. L. Lee, J. C. Woicik, [160] A. Paone, R. Sanjines, P. Jeanneret, H. J. Whitlow, E. Guibert,
J. H. Guo, J. D. Brock, D. A. Muller, D. A. Arena, D. G. Schlom, G. Guibert, F. Bussy, J. L. Scartezzini, A. Schüler, J. Alloys Compd.
L. F. J. Piper, Phys. Rev. B 2016, 94, 085105. 2015, 621, 206.
[133] N. B. Aetukuri, A. X. Gray, M. Drouard, M. Cossale, L. Gao, [161] K. A. Khan, G. A. Niklasson, C. G. Granqvist, J. Appl. Phys. 1988,
A. H. Reid, R. Kukreja, H. Ohldag, C. A. Jenkins, E. Arenholz, 64, 3327.
K. P. Roche, H. A. Dürr, M. G. Samant, S. S. P. Parkin, Nat. Phys. [162] H. Yoon, M. Choi, T. W. Lim, H. Kwon, K. Ihm, J. K. Kim, S. Y. Choi,
2013, 9, 661. J. Son, Nat. Mater. 2016, 15, 1113.
[134] M. M. Qazilbash, M. Brehm, G. O. Andreev, A. Frenzel, [163] A. Krammer, A. Gremaud, O. Bouvard, R. Sanjines, A. Schüler,
P. C. Ho, B. G. Chae, B. J. Kim, S. J. Yun, H. T. Kim, A. V. Balatsky, Surf. Interface Anal. 2016, 48, 440.
O. G. Shpyrko, M. B. Maple, F. Keilmann, D. N. Basov, Phys. Rev. [164] I. Balberg, B. Abeles, Y. Arie, Thin Solid Films 1974, 24, 307.
B: Condens. Matter Mater. Phys. 2009, 79, 075107. [165] J. Livage, F. Béteille, J. Livage, J. Sol-Gel Sci. Technol. 1998, 13, 65.
[135] J. Jian, A. Chen, Y. Chen, X. Zhang, H. Wang, Appl. Phys. Lett. 2017, [166] P. Jin, M. Tazawa, M. Ikeyama, S. Tanemura, K. Macák, X. Wang,
111, 153102. S. Olafsson, U. Helmersson, J. Vac. Sci. Technol., A 1999, 17, 1817.
[136] M. K. Liu, M. Wagner, E. Abreu, S. Kittiwatanakul, A. McLeod, [167] K. Shibuya, M. Kawasaki, Y. Tokura, Appl. Phys. Lett. 2010, 96,
Z. Fei, M. Goldflam, S. Dai, M. M. Fogler, J. Lu, S. A. Wolf, 022102.
R. D. Averitt, D. N. Basov, Phys. Rev. Lett. 2013, 111, 96602. [168] M. Zhu, H. Wang, B. Wang, H. Qi, D. Zhang, W. Lv, Ceram. Int.
[137] Y. Zhao, J. Hwan Lee, Y. Zhu, M. Nazari, C. Chen, H. Wang, 2019, 45, 1661.
A. Bernussi, M. Holtz, Z. Fan, J. Appl. Phys. 2012, 111, 053533. [169] I. G. Madiba, N. Émond, M. Chaker, F. T. Thema, S. I. Tadadjeu,
[138] G. Xu, P. Jin, M. Tazawa, K. Yoshimura, Appl. Surf. Sci. 2005, 244, U. Muller, P. Zolliker, A. Braun, L. Kotsedi, M. Maaza, Appl. Surf.
449. Sci. 2017, 411, 271.
[139] J. Li, J. Dho, J. Cryst. Growth 2014, 404, 84. [170] I. G. Madiba, N. Émond, M. Chaker, B. S. Khanyile, S. I. Tadadjeu,
[140] K. Nagashima, T. Yanagida, H. Tanaka, T. Kawai, Phys. Rev. B: Con- P. Zolliker, M. Izerrouken, N. Matinise, A. Braun, M. Nkosi,
dens. Matter Mater. Phys. 2006, 74, 13. M. Maaza, Nucl. Instrum. Methods Phys. Res., Sect. B 2019, 443, 25.
[141] H. Paik, J. A. Moyer, T. Spila, J. W. Tashman, J. A. Mundy, [171] A. Leone, A. M. Trione, F. Junga, IEEE Trans. Nucl. Sci. 1990, 37,
E. Freeman, N. Shukla, J. M. Lapano, R. Engel-Herbert, W. Zander, 1739.
J. Schubert, D. A. Muller, S. Datta, P. Schiffer, D. G. Schlom, Appl. [172] W. M. Tong, J. J. Yang, P. J. Kuekes, D. R. Stewart, R. S. Williams,
Phys. Lett. 2015, 107, 163101. E. DeIonno, E. E. King, S. C. Witczak, M. D. Looper, J. V. Osborn,
[142] L. L. Fan, S. Chen, Z. L. Luo, Q. H. Liu, Y. F. Wu, L. Song, D. X. Ji, IEEE Trans. Nucl. Sci. 2010, 57, 1640.
P. Wang, W. S. Chu, C. Gao, C. W. Zou, Z. Y. Wu, Nano Lett. 2014, [173] T. Yajima, Y. Ninomiya, T. Nishimura, A. Toriumi, Phys. Rev. B: Con-
14, 4036. dens. Matter Mater. Phys. 2015, 91, 205102.
[143] A. Gupta, J. Narayan, T. Dutta, Appl. Phys. Lett. 2010, 97, 151912. [174] L. Rodríguez, E. Del Corro, M. Conroy, K. Moore, F. Sandiumenge,
[144] D. Lee, J. Lee, K. Song, F. Xue, S. Y. Choi, Y. Ma, J. Podkaminer, N. Domingo, J. Santiso, G. Catalan, ACS Appl. Electron. Mater.
D. Liu, S. C. Liu, B. Chung, W. Fan, S. J. Cho, W. Zhou, J. Lee, 2020, 2, 1433.
L. Q. Chen, S. H. Oh, Z. Ma, C. B. Eom, Nano Lett. 2017, 17, 5614. [175] J. Cao, Y. Gu, W. Fan, L. Q. Chen, D. F. Ogletree, K. Chen,
[145] M. R. Bayati, R. Molaei, F. Wu, J. D. Budai, Y. Liu, R. J. Narayan, N. Tamura, M. Kunz, C. Barrett, J. Seidel, J. Wu, Nano Lett. 2010,
J. Narayan, Acta Mater. 2013, 61, 7805. 10, 2667.

Adv. Mater. 2023, 35, 2205294 2205294 (25 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

[176] P. Jin, S. Nakao, S. Tanemura, T. Bell, L. S. Wielunski, M. V. Swain, [207] B. Hu, Y. Ding, W. Chen, D. Kulkarni, Y. Shen, V. V. Tsukruk,
Thin Solid Films 1999, 343–344, 134. Z. L. Wang, Adv. Mater. 2010, 22, 5134.
[177] A. Rúa, R. Cabrera, H. Coy, E. Merced, N. Sepúlveda, [208] B. Fisher, J. Phys. C: Solid State Phys. 1975, 8, 2072.
F. E. Fernández, J. Appl. Phys. 2012, 111, 104502. [209] T. L. Wu, L. Whittaker, S. Banerjee, G. Sambandamurthy, Phys. Rev.
[178] H. Guo, K. Chen, Y. Oh, K. Wang, C. Dejoie, S. A. Syed Asif, B: Condens. Matter Mater. Phys. 2011, 83, 2.
O. L. Warren, Z. W. Shan, J. Wu, A. M. Minor, Nano Lett. 2011, 11, 3207. [210] Z. Zhao, J. Li, C. Ling, X. Zhao, Y. Zhao, H. Jin, Thin Solid Films
[179] D. Singh, B. Viswanath, J. Mater. Sci. 2017, 52, 5589. 2021, 725, 138643.
[180] J. Hay, B. Crawford, J. Mater. Res. 2011, 26, 727. [211] Z. Zou, Z. Zhang, J. Xu, Z. Yu, M. Cheng, R. Xiong, Z. Lu, Y. Liu,
[181] B. Viswanath, C. Ko, S. Ramanathan, Scr. Mater. 2011, 64, 490. J. Shi, J. Alloys Compd. 2019, 806, 310.
[182] F. Sandiumenge, L. Rodríguez, M. Pruneda, C. Magén, J. Santiso, [212] A. Bradicich, H. Clarke, E. J. Braham, A. Yano, D. Sellers,
G. Catalan, Adv. Mater. 2021, 33, 2004374. S. Banerjee, P. J. Shamberger, Adv. Electron. Mater. 2021, 8,
[183] J. O. Krisponeit, S. Fischer, S. Esser, V. Moshnyaga, T. Schmidt, 2100932.
L. F. J. Piper, J. I. Flege, J. Falta, Sci. Rep. 2020, 10, 22374. [213] K. Ito, K. Nishikawa, H. Iizuka, Appl. Phys. Lett. 2016, 108, 053507.
[184] Y. Yang, X. Mao, Y. Yao, H. Huang, Y. Lu, L. Luo, X. Zhang, G. Yin, [214] S. H. Bae, S. Lee, H. Koo, L. Lin, B. H. Jo, C. Park, Z. L. Wang, Adv.
T. Yang, X. Gao, J. Appl. Phys. 2019, 125, 082508. Mater. 2013, 25, 5098.
[185] H. T. Kim, B. G. Chae, D. H. Youn, G. Kim, K. Y. Kang, S. J. Lee, [215] Y. Jung, J. Jeong, Z. Qu, B. Cui, A. Khanda, S. S. P. Parkin,
K. Kim, Y. S. Lim, Appl. Phys. Lett. 2005, 86, 242101. J. K. S. Poon, Adv. Electron. Mater. 2021, 7, 2001142.
[186] A. Joushaghani, J. Jeong, S. Paradis, D. Alain, J. Stewart Aitchison, [216] S. Singh, T. A. Abtew, G. Horrocks, C. Kilcoyne, P. M. Marley,
J. K. S. Poon, Appl. Phys. Lett. 2014, 104, 221904. A. A. Stabile, S. Banerjee, P. Zhang, G. Sambandamurthy, Phys.
[187] T. Driscoll, J. Quinn, M. Di Ventra, D. N. Basov, G. Seo, Y. W. Lee, Rev. B 2016, 93, 125132.
H. T. Kim, D. R. Smith, Phys. Rev. B: Condens. Matter Mater. Phys. [217] B. Zhi, G. Gao, H. Xu, F. Chen, X. Tan, P. Chen, L. Wang, W. Wu,
2012, 86, 094203. ACS Appl. Mater. Interfaces 2014, 6, 4603.
[188] D. Ruzmetov, G. Gopalakrishnan, J. Deng, V. Narayanamurti, [218] H. Bian, Y. Y. Goh, Y. Liu, H. Ling, L. Xie, X. Liu, Adv. Mater. 2021,
S. Ramanathan, J. Appl. Phys. 2009, 106, 083702. 33, 1.
[189] D. Gu, H. Qin, X. Zhou, S. Xu, Y. Jiang, AIP Adv. 2018, 8, 015317. [219] X. Deng, S. Q. Wang, Y. X. Liu, N. Zhong, Y. H. He, H. Peng,
[190] S. B. Lee, K. Kim, J. S. Oh, B. Kahng, J. S. Lee, Appl. Phys. Lett. P. H. Xiang, C. G. Duan, Adv. Funct. Mater. 2021, 31, 2101099.
2013, 102, 063501. [220] Z. Tian, B. Xu, B. Hsu, L. Stan, Z. Yang, Y. Mei, Nano Lett. 2018,
[191] D. Li, A. A. Sharma, D. K. Gala, N. Shukla, H. Paik, S. Datta, 18, 3017.
D. G. Schlom, J. A. Bain, M. Skowronski, ACS Appl. Mater. Inter- [221] L. Shan, Z. Wang, L. Bao, S. Bao, Y. Qin, Y. Ling, G. Bai,
faces 2016, 8, 12908. J. Robertson, Y. Cai, R. Huang, Adv. Intell. Syst. 2022, 2100264,
[192] G. M. Liao, S. Chen, L. L. Fan, Y. L. Chen, X. Q. Wang, H. Ren, 2100264.
Z. M. Zhang, C. W. Zou, AIP Adv. 2016, 6, 045014. [222] W. Wei, T. Huang, S. Wang, W. Luo, T. Zhang, W. Hu, X. Chen,
[193] I. P. Radu, B. Govoreanu, S. Mertens, X. Shi, M. Cantoro, N. Dai, Phys. Status Solidi RRL 2019, 13, 1900383.
M. Schaekers, M. Jurczak, S. De Gendt, A. Stesmans, J. A. Kittl, [223] A. Joushaghani, J. Jeong, S. Paradis, D. Alain, J. Stewart Aitchison,
M. Heyns, K. Martens, Nanotechnology 2015, 26, 165202. J. K. S. Poon, Appl. Phys. Lett. 2014, 105, 231904.
[194] A. Zimmers, L. Aigouy, M. Mortier, A. Sharoni, S. Wang, [224] E. Goi, Q. Zhang, X. Chen, H. Luan, M. Gu, PhotoniX 2020, 1, 3.
K. G. West, J. G. Ramirez, I. K. Schuller, Phys. Rev. Lett. 2013, 110, [225] M. F. Jager, C. Ott, P. M. Kraus, C. J. Kaplan, W. Pouse,
056601. R. E. Marvel, R. F. Haglund, D. M. Neumark, S. R. Leone, Proc.
[195] B. Simon Mun, J. Yoon, S. K. Mo, K. Chen, N. Tamura, C. Dejoie, Natl. Acad. Sci. USA 2017, 114, 9558.
M. Kunz, Z. Liu, C. Park, K. Moon, H. Ju, Appl. Phys. Lett. 2013, [226] V. R. Morrison, R. P. Chatelain, K. L. Tiwari, A. Hendaoui,
103, 061902. A. Bruhács, M. Chaker, B. J. Siwick, Science 2014, 346, 445.
[196] J. Yoon, G. Lee, C. Park, B. S. Mun, H. Ju, Appl. Phys. Lett. 2014, [227] B. T. O’Callahan, A. C. Jones, J. Hyung Park, D. H. Cobden,
105, 083503. J. M. Atkin, M. B. Raschke, Nat. Commun. 2015, 6, 6849.
[197] S. Kumar, M. D. Pickett, J. P. Strachan, G. Gibson, Y. Nishi, [228] S. A. Dönges, O. Khatib, B. T. O’Callahan, J. M. Atkin, J. H. Park,
R. S. Williams, Adv. Mater. 2013, 25, 6128. D. Cobden, M. B. Raschke, Nano Lett. 2016, 16, 3029.
[198] K. Okimura, N. Ezreena, Y. Sasakawa, J. Sakai, Jpn. J. Appl. Phys. [229] K. A. Hallman, K. J. Miller, A. Baydin, S. M. Weiss, R. F. Haglund,
2009, 48, 065003. Adv. Opt. Mater. 2021, 9, 2001721.
[199] J. Kim, C. Ko, A. Frenzel, S. Ramanathan, J. E. Hoffman, Appl. [230] Y. Jung, H. Han, A. Sharma, J. Jeong, S. S. P. Parkin, J. K. S. Poon,
Phys. Lett. 2010, 96, 213106. ACS Photonics 2022, 9, 217.
[200] E. Freeman, G. Stone, N. Shukla, H. Paik, J. A. Moyer, Z. Cai, [231] T. Kang, Z. Ma, J. Qin, Z. Peng, W. Yang, T. Huang, S. Xian, S. Xia,
H. Wen, R. Engel-Herbert, D. G. Schlom, V. Gopalan, S. Datta, W. Yan, Y. Yang, Z. Sheng, J. Shen, C. Li, L. Deng, L. Bi, Nanopho-
Appl. Phys. Lett. 2013, 103, 263109. tonics 2020, 10, 909.
[201] J. Duchene, M. Terraillon, P. Pailly, G. Adam, Appl. Phys. Lett. 1971, [232] X. Zhou, L. Xhao, W. Zhen, Y. Lin, C. Wang, T. Pan, L. Li, G. Du,
19, 115. L. Lu, X. Cao, D. Li, Adv. Electron. Mater. 2021, 7, 2001254.
[202] J. Yoon, H. Kim, B. S. Mun, C. Park, H. Ju, J. Appl. Phys. 2016, 119, [233] H. Luo, B. Wang, E. Wang, X. Wang, Y. Sun, Q. Li, S. Fan,
124503. C. Cheng, K. Liu, Appl. Phys. Rev. 2019, 6, 041407.
[203] B. G. Chae, H. T. Kim, D. H. Youn, K. Y. Kang, Phys. B: Condens. [234] E. Corti, A. Khanna, K. Niang, J. Robertson, K. E. Moselund,
Matter 2005, 369, 76. B. Gotsmann, S. Datta, S. Karg, IEEE Electron Device Lett. 2020, 41,
[204] Y. Zhou, X. Chen, C. Ko, Z. Yang, C. Mouli, S. Ramanathan, IEEE 629.
Electron Device Lett. 2013, 34, 220. [235] R. Yuan, Q. Duan, P. J. Tiw, L. Ge, Z. Xiao, Z. Jing, K. Yang, C. Liu,
[205] J. Leroy, A. Crunteanu, A. Bessaudou, F. Cosset, C. Champeaux, C. Ge, R. Huang, Y. Yang, Nat. Commun. 2022, 13, 3973.
J. C. Orlianges, Appl. Phys. Lett. 2012, 100, 213507. [236] S. Kumar, R. S. Williams, Z. Wang, Nature 2020, 585, 518.
[206] W. Xue, G. Liu, Z. Zhong, Y. Dai, J. Shang, Y. Liu, H. Yang, X. Yi, [237] M. A. Nahmias, T. F. De Lima, A. N. Tait, H. T. Peng, B. J. Shastri,
H. Tan, L. Pan, S. Gao, J. Ding, X. H. Xu, R. W. Li, Adv. Mater. 2017, P. R. Prucnal, IEEE J. Sel. Top. Quantum Electron. 2020, 26,
29, 1702162. 7701518.

Adv. Mater. 2023, 35, 2205294 2205294 (26 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

[238] D. V Christensen, R. Dittmann, B. Linares-Barranco, A. Sebastian, [241] A. Parija, J. V. Handy, J. L. Andrews, J. Wu, L. Wangoh,
M. Le Gallo, A. Redaelli, S. Slesazeck, T. Mikolajick, S. Spiga, S. Singh, C. Jozwiak, A. Bostwick, E. Rotenberg, W. Yang,
S. Menzel, I. Valov, G. Milano, C. Ricciardi, S.-J. Liang, F. Miao, S. C. Fakra, M. Al-Hashimi, G. Sambandamurthy, L. F. J. Piper,
M. Lanza, T. J. Quill, S. T. Keene, A. Salleo, J. Grollier, D. Marković, R. S. Williams, D. Prendergast, S. Banerjee, Matter 2020, 2,
A. Mizrahi, P. Yao, J. J. Yang, G. Indiveri, J. P. Strachan, S. Datta, 1166.
E. Vianello, A. Valentian, J. Feldmann, et al., Neuromorphic [242] Y. Luo, S. Rezaei, D. Santos, Y. Zhang, J. V. Handy, L. Carrillo,
Comput. Eng. 2022, 2, 022501. B. J. Schultz, L. Gobbato, M. Pupucevski, K. Wiaderek,
[239] I. Bendixson, Acta Math. 1901, 24, 1. H. Charalambous, A. Yakovenko, M. Pharr, B.-X. Xu, S. Banerjee,
[240] J. V. Handy, W. Zaheer, A. R. M. Rothfuss, C. R. McGranahan, Proc. Natl. Acad. Sci. USA 2022, 119, 2115072119.
G. Agbeworvi, J. L. Andrews, K. E. García-Pedraza, J. D. Ponis, [243] J. L. Andrews, P. Stein, D. A. Santos, C. J. Chalker, L. R. De Jesus,
J. R. Ayala, Y. Ding, D. F. Watson, S. Banerjee, Chem. Mater. 2022, R. D. Davidson, M. A. Gross, M. Pharr, J. D. Batteas, B.-X. Xu,
34, 1439. S. Banerjee, 2020, Matter, 3, 1754.

Parker Schofield is a fifth-year Chemistry doctoral student advised by Sarbajit Banerjee at Texas
A&M University. He earned his undergraduate degrees in Chemistry and English from Utah State
University in 2018. His research interests pertain broadly to electrochemical energy storage, neu-
romorphic computing materials design, and doping in phase-transition materials.

Adelaide Bradicich is currently pursuing her Ph.D. in Materials Science and Engineering at Texas
A&M University. Under the supervision of Dr. Patrick Shamberger, her research focuses on the
development and characterization of vanadium dioxide devices, as well as modeling the electrical
and thermal behavior of that material under an electrical bias. She earned her B.S. in Chemistry
and Physics from Texas A&M University – Commerce.

Rebeca M. Gurrola is currently pursuing her Ph.D. in Materials Science and Engineering at Texas
A&M University. Under the supervision of Dr. Patrick Shamberger, her research focuses on syn-
thesis and characterization of vanadium dioxide thin film systems as well as defect engineering of
VO2 for neuromorphic applications. She earned her B.S. in Physics and B.A. in Mathematics from
St. Mary’s University.

Adv. Mater. 2023, 35, 2205294 2205294 (27 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

Yuwei Zhang received his Ph.D. degree in Mechanical Engineering at Texas A&M University
in 2021. He is currently a postdoctoral associate in the Department of Material Science and
Engineering at Texas A&M University. His research focuses on small-scale mechanical testing in
the field of energy storage and conversion.

Timothy D. Brown received his Ph.D. in Materials Science & Engineering at Texas A&M University
in 2019. Since then, he has worked with Prof. Stan Williams at Texas A&M and now with Suhas
Kumar at Sandia National Labs on the development and exploitation of Local Activity theory in
the design of neuromorphic and dynamical devices and systems. His research interests include
further development of Local Activity theory into simple and useful device engineering princi-
ples, codesign of the entire compute stack to implement energy-efficient “software-in-hardware”
paradigms, and investigation of new materials expressing nonlinear physics for neuromorphic
applications.

Matt Pharr is an associate professor and J. Mike Walker ‘66 Faculty Fellow in Mechanical
Engineering at Texas A&M University with a courtesy appointment in Materials Science &
Engineering. He serves as co-lead of Early-Stage Investigator Professional Development of
REMIND, an Energy Frontier Research Center focused on advancing the fundamental science
of neuromorphic computing funded by the United States Department of Energy. Recent distinc-
tions include an NSF CAREER Award, a TEES Young Faculty Fellow Award, an Association of
Former Students Distinguished Achievement in Teaching Award, and a TMS LMD Young Leaders
Professional Development Award.

Patrick J. Shamberger earned his Ph.D. in Materials Science and Engineering from the University
of Washington, Seattle, WA. Currently, he is an associate professor with the Department of
Materials Science and Engineering at Texas A&M University. Prior to this, he served as a
materials research engineer for the Air Force Research Lab in the Materials and Manufacturing
directorate. His current research focuses on the development of reversible phase transformations
for electronic and thermal applications. He is the recipient of several awards, including an NSF
Career award from the Metals and Metallic Nanostructures program.

Adv. Mater. 2023, 35, 2205294 2205294 (28 of 29) © 2022 Wiley-VCH GmbH
 15214095, 2023, 37, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adma.202205294 by Technion-Israel Institution Of Technology, Wiley Online Library on [08/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmat.de

Sarbajit Banerjee is the Davidson Chair Professor of Chemistry, Professor of Materials Science
and Engineering, and Chancellor EDGES Fellow at Texas A&M University. He serves as the
Executive Associate Director of REMIND, an Energy Frontier Research Center focused on
advancing the fundamental science of neuromorphic computing funded by the United States
Department of Energy. Recent distinctions include the TAMEST Edith and Peter O’Donnell Award
in Science and the American Chemical Society Stanley C. Israel Regional Award for Advancing
Diversity in the Chemical Sciences. He is a Fellow of the Royal Society of Chemistry and the
Institute of Physics.

Adv. Mater. 2023, 35, 2205294 2205294 (29 of 29) © 2022 Wiley-VCH GmbH