Articles

https://doi.org/10.1038/s41928-021-00631-8

Neural recording and stimulation using wireless

networks of microimplants
Jihun Lee 1, Vincent Leung 2, Ah-Hyoung Lee 1,3, Jiannan Huang4, Peter Asbeck4,
Patrick P. Mercier 4, Stephen Shellhammer5, Lawrence Larson1, Farah Laiwalla1 and
Arto Nurmikko 1,6 ✉

Multichannel electrophysiological sensors and stimulators—particularly those used to study the nervous system—are usually
based on monolithic microelectrode arrays. However, the architecture of such arrays limits flexibility in electrode placement
and scaling to a large number of nodes, especially across non-contiguous locations. Here we report wirelessly networked and
powered electronic microchips that can autonomously perform neural sensing and electrical microstimulation. The microchips,
which we term neurograins, have an ~1 GHz electromagnetic transcutaneous link to an external telecom hub, providing bidirec-
tional communication and control at the individual device level. To illustrate the potential of the approach, we show that 48 neu-
rograins can be individually addressed on a rat cortical surface and used for the acute recording of neural activity. Theoretical
calculations and experimental measurements show that the link configuration could potentially be scaled to 770 neurograins
using a customized time-division multiple access protocol.

D
etecting and stimulating physiological electrical activity at Furthermore, microdevices that provide electrical recording and
multiple points can provide insights into the operation of microstimulation in small animals have been demonstrated24–26.
targeted biological circuits such as those in the brain cor- A missing element is, however, an ability to network ensembles
tex. Current high-performance brain–machine cortical interfaces of microimplant nodes via a communication scheme that is scalable
are based on micromachined silicon-based ‘beds of needles’ or to large numbers. We recently proposed a wireless power and com-
electrocorticography (ECoG) planar electrode arrays connected munication approach, as well as mixed-signal integrated circuits,
to external active electronics, with tissue contact points number- for neural recording and stimulation, which could be used as build-
ing on the scale of hundreds1–6. Scaling up the number of channels ing blocks in the development of wirelessly networked microchip
to many thousands is expected to provide considerable perfor- ensembles27–30. In this Article, we report the development of a com-
mance benefits. However, achieving this with current technology plete wireless network of microimplants—termed neurograins—
creates challenges in terms of tethering architecture, data transfer and illustrate its capabilities via neural recording and stimulation
and implantability. One approach to scaling is to use the vertical in the rat cortex.
dimension, with multiple multiplexed probe sites microfabricated
along rigid penetrating silicon shanks (for example, the Neuropixels Wireless neurograin system
approach7). Techniques for high-throughput electrode implanta- Figure 1 summarizes the main features of the neurograin sys-
tion along with wireless brain implants are also being explored (for tem, reported here for epicortical ECoG recording use, where
example, the Neuralink approach8), in addition to wireless methods wireless-communication-compatible microcircuits (application-
that entirely eliminate the tethering of monolithic implants to exter- specific integrated circuits (ASICs)) are realized in a 65 nm RF
nal electronics9–16. complementary metal–oxide semiconductor (CMOS) process, each
An alternative approach to fabricate microscale sensor archi- chiplet occupying a total volume of ~0.1 mm3 (Fig. 1b). The electro-
tectures is to use multichannel implant ensembles that consist of magnetic interface for communication and near-field power har-
individual stand-alone microdevices. For example, piezoelectric vesting between an external hub and the ensemble of neurograins is
materials have been integrated into implants for energy harvesting conceptualized in Fig. 1a, which also shows the additional subcuta-
and data transmission using ultrasound as the intermediary modal- neous resonant co-planar relay coil designed to improve the WPT
ity17; the neural signal recording and stimulation capabilities of these efficiency. A block-diagram-level description of the key electronic
devices have also been tested in vivo18,19. Purely electromagnetic functions including the dedicated telemetry blocks is shown in
wireless power transfer (WPT) and communication in the near field Fig. 1c for the recording and stimulating versions of the
(inductive) have also been studied in millimetre-sized devices20–22. neurograin system.
Achieving good electromagnetic coupling is inherently challeng- Standard amplitude-shift keying (ASK), frequency-shift keying,
ing for small-area chip-scale antennas17,23, but radio-frequency load-shift keying or infrared–ultrawideband modulation schemes
(RF) techniques can—if properly optimized—provide an efficient have been used to communicate with microimplants, demonstrated
approach for wireless energy harvesting and telecommunication. so far in limited channel bandwidth systems, typically below 1 Mbps

School of Engineering, Brown University, Providence, RI, USA. 2Electrical and Computer Engineering, Baylor University, Waco, TX, USA. 3Graduate
1

School of Convergence Science and Technology, Seoul National University, Seoul, South Korea. 4Electrical and Computer Engineering, University of
California San Diego, San Diego, CA, USA. 5Qualcomm Inc., San Diego, CA, USA. 6Carney Institute for Brain Science, Brown University, Providence, RI, USA.
✉e-mail: [email protected]

Nature Electronics | www.nature.com/natureelectronics

Articles Nature Electronics

a b
Wireless Tx coil RF power/downlink 2 mm
RF link
Uplink

Coaxial
cable

Cortex
Neurograin
Portable
transceiver Relay coil 10 mm
1 mm

c
Tissue (for
example, skin, Power
management Amp
External wireless hub skull, dura)

Chip controller
On-chip coil
Neurograin
recording
Software-defined radio BPSK modulation
Uplink Logic ‘0’ Logic ‘1’
(10 Mbps)
FPGA ADC
Amp
Transceiver

Relay coil
Tx coil

Power
management

Chip controller
On-chip coil
stimulation
Neurograin

Processor
Duplexer
Power/
downlink ASK-PWM demodulation
(1 Mbps) Logic ‘1’ Logic ‘0’
2T T T 2T

To PC
Neurograin
array

Fig. 1 | Wireless neurograin system for distributed autonomous networking. a, Concept of a transcutaneous RF power and data link for a neurograin array.
b, Handful of chiplets (size, 650 μm × 650 μm × 250 μm) present on a United States dime (the inset shows an optional post-process-integrated microwire for
intracortical access, which has been used to demonstrate microstimulation in this paper). c, Block diagram of the overall electronic system, consisting of the
external RF hub, inductive coupling link and key features of the neurograins for recording and stimulation, respectively. FPGA, field-programmable gate array;
Amp, amplifier; PC, personal computer; T, time interval.

(refs. 25,31,32). For large ensembles of spatially distributed implants, a stimulation site operates as a microminiaturized autonomous sensor
more versatile networking protocol with larger data bandwidth is node with its own independent wireless telemetry channel. For com-
needed. Multichannel techniques such as frequency-division mul- parison with the work presented here, Supplementary Table 1 lists
tiple access (FDMA) and code-division multiple access (CDMA) examples of other state-of-the-art microdevices.
are well established in cellular networks33. However, when consid-
ering their adaptation to microimplants, the FDMA approach can System-on-chip microimplant
encounter channel size restrictions as well as challenges with indi- Individual microchips operating as a part of wirelessly networked
vidual band tuning34. CDMA implementation requires fine timing ensembles need to incorporate multiple functions from energy
calibration and has been demonstrated only as a two-channel sys- harvesting to bidirectional data communication as well as neural
tem so far35 (Supplementary Information provides a more detailed recording and/or stimulation. The neurograins shown in Fig. 1
comparison in ‘Comparison of possible TDMA, FDMA and CDMA are submillimetre-sized system-on-chip (SoC) systems designed
approaches for neurograins’). as ASICs, which must be tightly constrained for ultralow-power
In light of these considerations, in our neural implants, we have operation. In this work, we have designed ASICs using the TSMC
opted to implement a specific power-efficient time-division multiple 65 nm mixed-signal/RF low-power CMOS foundry process, with
access (TDMA) network strategy for communication and the active no off-chip components. The photomicrograph shown in Fig. 2a
management of system latency, which provides the necessary scalabil- indicates the overall footprint and main functional circuit areas of a
ity as well as power and size efficiency in our neurograin multi-user neurograin chip specifically fabricated for electrical microstimula-
network. The key accomplishment in this paper is the integration of tion according to the block diagram shown in Fig. 2b. The corre-
ECoG-type neural recording, intracortical stimulation and telecom- sponding block diagram and footprint for a recording neurograin
munication functions on a single submillimetre microchip and the are shown in Supplementary Fig. 1.
demonstration of a multinode wireless microsensor system includ- Both recording and stimulating the neurograin chiplets share a
ing the initial application in vivo. In this approach, each recording or common ‘RF interface’ (microcircuits indicated by the blue arrow

Nature Electronics | www.nature.com/natureelectronics

Nature Electronics Articles
a b c
10 dB/div Ref –11.79 dBm
Three-turn coil Common Chip variants Log

–21.8

VDD Comp
ASK- –31.8
Tx tone
Stim BPSK data
PWM Downlink
–41.8

Pad 1 Pad 2

Rectifier
data
–51.8

Osc Rectifier Clk

–61.8

Osc Addr –71.8

Device ID Digital –81.8

Uplink logic LFSR –91.8

Vsw
Downlink Mod data Stim –102

decoder 100 µm Center 907.22 MHz Span 100.0 MHz

Res BW 910 kHz VBW 910 kHz Sweep 1.00 ms (1,001 pts)

d 0 e 0.4 f
Duplexer
0.2
density (dB Hz–1)

–20
Power spectral

Voltage (V)
0
–40
PA
–0.2
–60 SDR
–0.4 To coil
100 101 102 0 1 2 2 cm
Frequency (Hz) Time (ms)

Fig. 2 | Design of neurograin chip and wireless power and data link. a, A photomicrograph of the stimulating neurograin (outer dimensions, 500 μm × 500 μm)
comprising a three-turn microcoil (identical for each neurograin), rectifier, downlink decoder and device ID. A pair of contact electrode pads forming the
physiological interface with tissue is also identified. b, Block diagram of the ASIC circuit. The blue arrow highlights the RF interface (circuits common to the
neurograins of different modalities), and the red arrow points to blocks dedicated to particular physiological interfaces (recording, stimulation, and so on).
Stim, stimulation; VDD, voltage supply; VSW, switching voltage; Osc, oscillator; Clk, clock; Comp, comparator; Addr, address; LFSR, linear-feedback shift register.
c, RF spectrum showing the baseband Tx tone and BPSK-modulated backscattered data lobe of the receiving channel. d, Normalized power spectral density of
a recording neurograin detecting 40 Hz, 2 mV peak-to-peak sinusoids injected into the surrounding saline. e, Biphasic current stimulation voltage waveforms
generated by the stimulating neurograin for three different pulse widths per phase (100, 200 and 400 μs). A 20 kΩ load was used. f, Photograph of the
benchtop external RF telecom hub consisting of a software-defined radio (SDR), a power amplifier (PA) and a duplexer.

in Fig. 2b, while the red arrow indicates the purpose-specific sub- Supplementary Fig. 3a. The network architecture and data rates
circuit for microstimulation in this case). This common RF inter- highlighted herein allow the incorporation of downlink commu-
face includes an efficient power-harvesting circuit feeding off a nication around the Tx tone to complete a real-time bidirectional
three-turn on-chip coil (Fig. 2b) and utilizing a three-stage, low communication loop, as described in the next section. For neural
turn-on, deep n-well n-type metal–oxide–semiconductor transistor stimulation applications, we choose a 1 Mbps downlink rate, suf-
rectifier circuit with a low transistor threshold36. Up to 48.5% recti- ficient to include a range of command signals for commonly used
fication efficiency at −10 dBm input is achieved—a critical feature electrical microstimulation protocols for the mammalian brain.
for the feasibility of neurograin microimplants—through the avoid- A particular challenge for designing a communication proto-
ance of body effects in the n-type metal–oxide–semiconductor col for ensembles of neurograins is the inherent variance in the
transistors in this design37. An RF backscattering data uplink chan- on-chip-generated clock frequency and phase, which arises from the
nel operates at a rate of 10 Mbps near the ~1 GHz baseband wireless finite variation in the power harvested by each chip. This variation
powering link and is enabled through a modulator (‘Mod’ in Fig. 2b). makes synchronization difficult to achieve under usual protocols.
The specific choice of wireless frequency near 1 GHz, based on We tackled this challenge by using a custom ASK and pulse width
electromagnetic simulations, represents our compromise between modulation (ASK-PWM) scheme to implement the 1 Mbps down-
trade-offs to achieve an acceptable cross-section of the near-field link, as demonstrated by the circuit shown in Fig. 2b. This modula-
microantenna while minimizing the penalty by absorptive losses tion approach provides network synchronicity and is robust across
due to body tissue (‘The rationale for the 1 GHz range frequency the measured range of chip clock rates. The ASK-PWM approach,
selection’ in Supplementary Information). which is compatible with our proposed networking architecture
We use binary phase-shift keying (BPSK) modulation for uplink (outlined in the next section), encodes each bit as a combination
communications. The modulator operates by toggling a tuning of high- and low-power (amplitude) pulses of variable duration.
capacitor to enable RF backscattering. While ensuring that the mag- The received data on individual neurograins may thus be decoded
nitude of its input impedance is maintained at a constant value, the using only the relative durations between the high and low states,
phase modulation is maximized by selecting the value of the tog- independent of the internal clock frequency. Supplementary Fig. 2
gling capacitor to maintain a stable power supply to the chip37. Data illustrates the principle of the ASK-PWM scheme and the decoding
are upconverted before backscatter using an on-chip, free-running circuit implemented in this work.
relaxation oscillator with an operational frequency near 30 MHz; A unique device identifier (ID) is imperative for the functional
the latter enables spectral isolation of the BPSK-modulated signal implementation of an addressable wireless network. For the neu-
from the transmitted RF powering (Tx) tone, as shown in Fig. 2c. rograin microimplants, we have successfully tested two different
High-fidelity isolation of the backscattered data facilitates their ID approaches: (1) a physical digital address, engraved on-chip by
low-noise demodulation and recovery at the external hub (Fig. 2f) laser ablation38, and (2) an addressing scheme based on physically
for online or offline analysis (Methods). An example of the demod- unclonable functions (PUF)39,40; the latter approach is discussed
ulation of the BPSK-modulated backscattered signal is shown in below in a neural recording example. In the case of the laser ablation

Nature Electronics | www.nature.com/natureelectronics

Articles Nature Electronics

approach (an example is demonstrated below for microstimulation and transmitted over 82 µs (10 Mbps rate) in a pseudo-randomly
(‘device ID’ in Fig. 2a)), a user-selectable 10-bit address is physi- determined time slot. This hardware-efficient approach, however,
cally imprinted onto an array of ten metal interconnect traces on the has a network capacity for ~100 nodes, beyond which the possibility
top aluminium metal layer of the chip by selectively ablating a spe- of packet collisions (as shown in Supplementary Fig. 4a) increases
cific number of traces with a 532 nm pulsed laser. By contrast, the for an ensemble of free-floating microimplants due to unscheduled
address circuits using the PUF approach take advantage of intrin- backscattering and chip-clock frequency variations.
sic 65 nm CMOS fabrication process variations to generate unique, In principle, an externally commanded, programmable, synchro-
randomized bit patterns. The number of bits in a PUF address has to nized ‘call-and-respond’ network offers a considerable enhance-
be carefully considered to minimize the chance for address overlap ment in both network efficiency and adaptive channel selectivity
for a given population of networked nodes. In our case, we have compared with ‘autonomous’ operation, but at the cost of integrat-
implemented 16–20-bit PUFs, anticipating an insignificant address ing downlink communication circuits on the microdevices. In an
collision rate for ensembles of up to 1,000 nodes. In general, the exploratory investigation of this architecture, integrated onto the
PUF approach offers distinct advantages in scalability due to the neurograin SoCs, we have found a capability to bidirectionally com-
compact footprint and absence of labour-intensive post-processing; municate with 425 microdevices within a 100 ms period, theoreti-
however, the unique address needs to be ‘discovered’ for every neu- cally extendable up to 770 microdevices through the integration of
rograin. For the discovery of the PUF address, each neurograin has a shorter ‘call’ downlink (ongoing research), which would be the
to be individually tested before the experiments as they transmit likely eventual choice for truly large-scale multinode communica-
their PUF address through the uplink. tion for distributed microsensor networks. A detailed discussion
The front end of a neurograin SoC integrates high-fidelity physi- for TDMA scalability is provided in Supplementary Information
ological sensing (recording) and actuating (stimulation) circuits with (‘Scalability of the neurograin TDMA network’). To validate and
a special focus on ultraminiaturization and extremely low power compare the two communication protocols, benchtop experiments
operation. To this end, we have designed an analogue front end (AFE) were performed in air using a three-coil wireless system with 8 mm
for recording, a version of which is shown here for the acquisition separation between the Tx and relay coil.
of ECoG signals using a d.c.-coupled voltage-to-current converter The ‘autonomous’ TDMA protocol relies on the chip’s unique
merged with an analogue-to-digital converter via an area-efficient device ID, and as the chip only uses the uplink in its transmitting
capacitor-less approach41. The AFE directly senses neural poten- path, this built-in prefabricated PUF ID probabilistically determines
tials (without the need for bulky a.c.-coupling capacitors) with the device’s initial communication time slot in the network queue.
input-referred noise of 2.2 µV and 500 Hz bandwidth, and it occupies Data are periodically backscattered at the predesigned network
an overall footprint of 100 μm × 100 μm while recording at a 1 kHz latency (nominally, 100 ms). The on-chip clocks are designed to be
sampling rate and 8-bit resolution. Benchtop characterization mea- free running at 30 MHz with a variation of ±12.3%; this variation is
surements are summarized in Fig. 2d and Supplementary Fig. 3b. due to the spatial dependence of the magnetic coupling strength and
Electrical microstimulation, on the other hand, is facilitated the resulting variation in the received RF power. Figure 3b shows the
through circuits designed to deliver biphasic current pulses with measured transient data received at the external hub from a network
programmable pulse widths (here 100, 200 and 400 μs) as com- of 64 chips. The ‘pulses’ indicate periodic data packets sent by the
manded by the external wireless hub through the downlink (Fig. nodes. For each data packet, the decoded bits are compared with the
2e,f). At present, the maximum level of current stimulation depends predefined sequence to determine the bit error rate (BER). Due to
on the chip power level and can range up to 25 μA, as shown in the randomized time slots, among the 1,160 data packets detected,
Supplementary Fig. 3c; future work will incorporate stimulation 128 packets showed effects by packet collisions, that is, the simulta-
amplitude regulation across the network. We implemented pro- neous transmission of data, which checks with the simulation of the
grammable stimulation control to include both single-shot and packet success rate (no collisions) shown in Supplementary Fig. 5b.
100 Hz pulse train paradigms, the latter mitigating the data bur- Since a collision only affects portions of a given packet, demodula-
den on the downlink. The single-shot mode is capable of trigger- tion of the remainder of the data within the collided packets was
ing burst stimulation at rates faster than 500 Hz using the recursive still possible, where data was recovered with an average BER of
downlink, which is above the frequency commonly used for electri- 1.82%. When entirely excluding the collided packets, the average
cal microstimulation in the cortex. BER was 0.007%, showing that packet collision was the dominant
factor for bit errors in autonomous TDMA. The signal-to-noise
Networking of neurograin ensembles ratio (SNR) and BER for 1,160 packets are shown in Supplementary
We employed a TDMA-based networking architecture using a sin- Fig. 4b. An example of the characterization of the long-term stabil-
gle carrier frequency to enable periodic, scheduled communication ity of WPT and data telemetry scheme is shown in Fig. 3c. When
to and from the ensembles of implanted devices. The efficient use 1,800 data packets are subsampled from 64 neurograins over 30 min,
of finite channel bandwidth while maximizing the networked node the spread of their relative pulse powers mapped the harvested vari-
count is of utmost importance in a neural prosthesis application, able energy as a function of the chip’s spatial location. Packet RF
which also demands low-latency data-streaming capabilities. In this power and recovered chip clock frequency are plotted in Fig. 3c,
paper, we compare two different TDMA-based schemes: ‘autono- with each colour representing an individual neurograin, while the
mous’ and ‘call-and-respond’ approach. k-mean clustering into 64 groups shows a well-isolated, stable clock
The key trade-off in large-scale network design is the balance frequency over time.
between hardware complexity and networking efficiency: the former As noted, to avoid packet collision inherent in the autonomous
defines stringent power and area constraints for the microdevices, TDMA mode, we also explored a ‘call-and-respond’ TDMA net-
while the latter is critical for scalability. In recognition of these con- work, where the ASK-PWM protocol for downlink communication
siderations, the neurograin SoC network was initially realized using is integrated in the microdevices. The unique device IDs (16-bit PUF
unidirectional ‘autonomous’ TDMA since this is a simple, compact address), implemented on-chip, are recovered from the ensemble
and low-power approach for early ensemble testing at the benchtop through an automatic start-up communication. The IDs are subse-
and in rodents where channel counts are limited by the anatomy. quently used to externally command individual devices through the
In the autonomous mode, there is no synchronization between the ASK-PWM link. Each downlink command is received and down-
nodes. Each microdevice needs to buffer and store up to 100 ms of link codes are deciphered across the ensemble in a synchronized
data (820 bits, PUF address and recording); these data are packetized manner. If the received downlink code matches a given device’s

Nature Electronics | www.nature.com/natureelectronics

Nature Electronics Articles
a b c
1.0 32

0.8

Clock frequency (MHz)

30

Amplitude (a.u.)
0.6
28
0.4

26
0.2

0 24
0 20 40 60 80 100 1.0 1.5 2.0
Time (ms) Backscattering amplitude (a.u.)
d e f
DL UL DL UL DL UL 12
BER
1.0 10–1 SNR 11

10
0.8

Mean BER (%)

Amplitude (a.u.)

SNR (dB)
9
0.6
8
0.4 –2
10 7
0.2 6

0 5
0 0.2 0.4 0.6 0.8 0 10 20 30
Time (ms) Number of chips

Fig. 3 | Data communication demonstration on the benchtop with 8 mm separation between the Tx and relay coil in air. a, Sixty-four autonomous TDMA
chips, with placement within a relay coil having an area of 20.4 mm × 20.4 mm. b, A transient recording on the uplink data packets from multiple chips
under ‘autonomous’ TDMA. c, Recovered clock frequencies and backscattering amplitudes of 1,800 data packets sampled for over 30 min from 64 chips.
d, A 32-node call-and-respond network within a relay coil. e, A transient recording on the downlink (DL) and uplink (UL) under ‘call-and-respond’ TDMA.
f, Average BER from 1,000 uplinked bits versus data packet SNR of 947 data packets from 32 chips individually identified by PUF addresses.

address, it transmits a data packet (‘respond’). Figure 3e shows a This approach is not only able to achieve electrode impedances suit-
snapshot of the transient received signal at the external hub with a able for epicortical ECoG recording (Fig. 4b, left) but also suitable
strong downlink call of 40 μs and weak, uplink backscattered data for the attachment of microwires (a 50-μm-diameter tungsten wire
of 100 μs for a network of 32 neurograins. Successful communica- (Fig. 4b, right)) for use as intracortical electrodes such as that in
tions, measured by the response rate to the downlink, occur with an the microstimulation experiments below. For long-term chronic
average efficacy of 94.7% across this network, and we were also able device implantation (not the goal of this paper), we have reported
to identify three other devices with <90% efficacy (Supplementary elsewhere the development of ultrathin dielectric coatings, applied
Fig. 7a). This can be possibly due to CMOS fabrication and dic- by atomic vapour deposition, to obtain biocompatible, hermetically
ing process variations since all the available chips were tested and sealed neurograins42.
no prior screening was done. The uplinked 967 data packets from While each modality (recording and stimulation) is anticipated
these chips, comprising 1,000-bit predefined sequences, yielded an to ultimately integrate into a single, bidirectional TDMA link, we are
average BER of 0.004%, as shown in Fig. 3f (Supplementary Fig. 7c currently able to leverage this protocol to fulfil the specific needs of
for each packet), validating the ability of the downlink/uplink and each separate function across neurograin ensembles. The recording
the network to withstand variations in SNR and on-chip power. chip reported here generates the uplink under autonomous TDMA
Further work on the ‘call-and-respond’ approach is subject to ongo- (see the previous section). Since the head/skull size of a rat and the
ing research. associated craniotomy limits the number of practically implantable
neurograins (here N = 48), the probability of data packet collision is
Wireless recording and microstimulation by neurograins low. As examples of recordings in saline (Supplementary Fig. 8a),
We have built and assessed a fully functional system (as shown in various current waveforms were injected from a nearby microwire
Fig. 1), both on a benchtop (saline solution) and in an in vivo rat either as proxy neural signals (Fig. 4c) or sinusoids at 200 and 40 Hz
model for neural recording as well as electrical microstimulation, (Supplementary Fig. 8b,c), and the recordings by the neurograins
focusing on the ‘autonomous’ TDMA network implementation. For are qualitatively compared with those acquired by a commercial
multichannel, ensemble neural recording, the external wireless hub wired neural recording system (using the multichannel neural
receives neural data as TDMA digital streams. In the stimulation signal processor ‘Nomad’ by Ripple Neuro). A quantitative com-
mode, the hub coordinates the ASK-PWM downlink for space– parison is difficult with a large rack-mounted commercial system,
time-specific microstimulation. but the results show how further improvement is needed for neu-
Converting fabricated microchips into functional microsensors/ rograins (recorded signal quality) even if they capture the low- and
stimulators requires further post-processing after their dicing from high-frequency feature of local field potentials quite well.
the wafer. For a fast turnaround, we have developed a polymer-based In the initial in vivo assessment, we were able to deploy neu-
microfabrication technique38 to deposit planar gold microelectrodes rograin ensembles while identifying 48 unique PUF addresses
(60 μm pad size) atop the aluminium input contact pads of the chip. (‘In vivo data analysis: role of the PUF address’ in Supplementary

Nature Electronics | www.nature.com/natureelectronics

Articles Nature Electronics

a b
Neurograin
array

Tx coil
100 µm 1 mm

c d e 400
Neurograin 1

Voltage (µV)
1.0 Neurograins 125
200
0 0
0.5
Voltage (mV)

Neurograin 1
–200 Ketamine
–100
0 0 200 400 600 800 1,000
75 Neurograin 2
Time (ms)

Voltage (µV)
–0.5 0
200
–1.0 –75 0
–200 Isoflurane
Voltage (µV)

1 2 3 4 5 6 Neurograin 3
100
Time (s) 0 200 400 600 800 1,000
0 Time (ms)

Voltage (µV)
Reference –100 400
200
Neurograin 4 0
Voltage (mV)

2 100

Neurograin 2
–200
–400 Ketamine
0
0 0 200 400 600 800 1,000
–100 Time (ms)
Neurograin 5

Voltage (µV)
125
–2 200
0 0
–200 Isoflurane
1 2 3 4 5 6 –175
Time (s) 1s 0 200 400 600 800 1,000
Time (ms)

f g h Chip a
Downlink 180 180 Stim
Time (ms) Chip b
0.5
Firing rate (spikes per second)
Ch 1

0 160 160 100

LFP amplitude (µV)

LFP amplitude (µV)

–0.5
0 0.2 0.4 0.6 0.8 1.0 1.2 140 140 80
Voltage (mV)

0.5
Ch 16

0 120 120 60
–0.5
100 100 40
0 0.2 0.4 0.6 0.8 1.0 1.2
0.5 80 80
Ch 44

0 20
–0.5 60 60 0
0 0.2 0.4 0.6 0.8 1.0 1.2 200 400 200 400 200 400 600
Time (s)
Pulse width (µs) Pulse width (µs) Time (ms)

Fig. 4 | Neurograin recording and stimulation in saline and in vivo acute rat model. a, Photographs of the in vivo implant and head-mounted components
(left) and an array of recording neurograin encapsulated by polydimethylsiloxane (PDMS) along with the relay coil on a polyimide board (right). b,
Photomicrographs of gold electrodes fabricated on a chip (left) and microstimulator with intracortical tungsten electrodes (right). c, Injected ECoG signal
recordings from neurograins (top) and a commercial device (Ripple Neuro ‘Nomad’) (bottom) in saline. Distance difference between the signal source and
the recording site resulted in a difference in the signal amplitude. The Nomad system used a distal ground electrode, whereas neurograins measured the
differential between the two on-chip gold electrodes. d, Multichannel neurograin recordings on a spontaneous low-frequency oscillatory wave from the
anaesthetized rat. e, Recorded raw signals on 5 Hz electrical stimulation (blue dots) and the evoked neural response under ketamine and 3% isoflurane
using a recording neurograin array. Stimulation was intracortically delivered while the ECoG recording was performed in the epicortical region. f, Electrical
stimulation trains at 100 Hz from the stimulated neurograin array in saline with downlink artefacts indicated by blue arrows. g, Amplitudes of post-stimulus
LFP response depending on the stimulation pulse width (per phase) from two chips (left and right). Up to 25 μA current was delivered to the tissue from
each device. Error bars are presented as mean values ± s.d. (each consisting of 40 trials). h, Sum of the firing rate from three recording channels before and
after 400 Hz burst stimulation from a neurograin, showing increased post-stimulus firing rates (averaged over five trials with a bin size of 10 ms).

Information and results shown in Supplementary Fig. 9). The skin thickness of a rat (a maximum of 2.5 mm for each tissue layer),
unique address on each chip ensures continuously identifiable the associated RF attenuation is of the order of 2.5 dB, a minor effect
neural data flow from each recording neurograin, showing that on total losses (Supplementary Fig. 15d).
communication with 48 channels was established. For the in vivo The geometry for testing the neurograin system in the in vivo
experiments, the skull was not replaced, in part because of the large rat model (acute) involved a six-channel commercial intracortical
craniotomy required to accommodate at least 48 neurograins made stimulation device using tungsten microwire electrodes, while the
this impractical in acute experiments. However, given the skull and epicortical neurograins recorded the evoked neural activity (‘In vivo

Nature Electronics | www.nature.com/natureelectronics

Nature Electronics Articles
experiment setup for neurograin’ in Supplementary Information neurograin implants, we have utilized a three-coil architecture for
provides a detailed description). The stimulation pulse width was inductive coupling (Methods). Design constraints for this coil sys-
set at 1 ms per phase and a current up to 50 μA was delivered. An tem—comprising the external Tx coil, implanted Tx-size-matched
array of recording neurograins was placed on the animal’s cortex, relay coil and on-chip receiver (Rx) microcoils—are informed by
covering most of the motor and sensory areas in proximity to the high-performance electromagnetic simulations. The Tx and relay
implanted wired stimulation device, and the external transmission coils were both designed in a geometry that can be viewed as a
coil (Tx) was positioned above the animal’s head (Fig. 4a). As shown de facto superposition of unit subcoils, capable of linear or other
in Fig. 4d and Supplementary Fig. 10a,b, in the absence of an exter- geometrical area extension for tailored areal coverage while main-
nal stimulus, the neurograin devices were able to record spontane- taining WPT efficiency, as shown in Fig. 5a–c. Leveraging the
ous ECoG signals, prominently capturing the subhertz slow-wave same underlying electromagnetic principles, we have been able to
oscillations characteristic of ketamine anaesthesia43. In the current custom design the Tx/relay coil geometries to tailor the different
experiments, up to 12 channels were able to record low-noise sponta- versions of the system to optimally accommodate the anatomi-
neous cortical activity signals, while others showed higher noise lev- cal structure of both future primate experiments and the present
els due to possible combinations of neurograin location issues, less rodent model (dimensions and simulation parameters are shown in
than perfect electrode metal/tissue interfaces and so on. However, Supplementary Table 2). In each case, the characteristic coil layout
when the cortex was actively stimulated using 5 Hz current pulses, generates current flow patterns (Supplementary Fig. 13a,b) where
the neurograin ensemble captured clear post-stimulus-evoked field the respective directions of magnetic flux of the adjacent subunit
responses (not hampered by stimulation artefacts). The results of coils are opposite to each other, as shown in Fig. 5d. This design
such raw recordings are shown in Fig. 4e. Superimposed traces permits a lateral expansion of the power-harvesting area, without
shown in Supplementary Fig. 10c from multiple trials show that the deteriorating the critical quality factors of the coil or coupling fac-
recorded evoked responses persist for 20–70 ms after a stimulation tors between the coils. The on-chip microantenna design (Rx) is
event. In contrast, recorded responses under isoflurane anaesthe- optimized as an Rx coil using a thick aluminium top metal layer
sia (which inhibits cortical activity) shows conspicuously smaller available within our semiconductor foundry process. The Rx coil
amplitudes, thereby eliminating the likelihood of the recorded sig- was optimized to be a three-turn design since for a lower number of
nal originating from spurious internal noise or stimulus artefacts. turns, coupling to the high load impedance gets less efficient, while
To assess the stimulating version of neurograins, we demon- for a higher number of turns, self-resonance decreases the qual-
strate here the laser-inscribed address approach where we sched- ity factor of the microantenna. The coil has an outer diameter of
uled the delivery of electrical stimulation parameters for each target 500 μm, and is able to accumulate magnetic flux with an efficiency
chip through the use of the TDMA ASK-PWM downlink. Figure compatible with powering the on-chip operations of the high-load/
4f demonstrates the programmed, patterned delivery of electrical low-power ASIC at 1 GHz (Fig. 2a).
stimulation pulses across a neurograin array; RF modulation by One way to quantitatively analyse the electromagnetics of the
the downlink communication manifests as an ‘artefact’ (used here wireless link in situ—without resorting to the confounding wired
for timing reference). We have verified that ensembles of chips are measurements due to external probe size—is to measure the clock
individually responsive to the downlink commands and are able to frequency of the chip, which serves as an estimate of the power
deliver electrical microstimulation in a spatially distributed manner. level received by the chip, as per the simulated frequency–power
Supplementary Fig. 11a shows the downlink-coordinated stimula- relationship shown in Supplementary Fig. 3d. We have used a
tion trains from a population of 12 neurograins, each reliably fol- liquid-form head phantom construct as a proxy for a physiologically
lowing its downlink command (over three cycles in this case). relevant RF environment, and analysed the effect of tissue path loss
For the acute rat experiments with ensembles of stimulating and surrounding tissue permittivity on the WPT and data link47. We
neurograins, we microfabricated the auxiliary device (as shown in mapped the WPT efficiency at 64 different spatial positions inside
Supplementary Fig. 11b,c), where an array of conventional recording a primate model relay coil using a location sampling grid with an
microwires was integrated into the same polymer holder platform x–y pitch of 1.0 mm (Fig. 5e), and in a rodent model coil at 26 posi-
as the wireless stimulating microdevices, connected to commer- tions with an x–y pitch of 1.0 mm and 0.8 mm (Fig. 5f). Based on
cial multichannel (wired) recording electronics. This hybrid device the comparative anatomical characteristics, the separation between
enabled a detailed validation of the physiologically relevant neural the Tx coil to relay coil was 8 and 5 mm in the primate and rodent
activation on the operation of neurograin stimulation. As an exam- model coil, respectively.
ple, we targeted a single neurograin in the implanted ensemble, As expected, the Rx coil achieves its highest efficiency when in
and programmed it to inject a 100 Hz intracortical stimulus pulse close spatial proximity to the relay coil perimeter, and the coupling
train (up to 25 μA current). We found a number of recording chan- decreases when moving towards the centre of each subunit coil
nels picking up post-stimulus local field potential (LFP) responses (Fig. 5e,f). For the primate model coil, the WPT efficiency ranges
evoked by a single neurograin, reflecting extended cortical network from −36.77 dB to −30.73 dB (centre to edge). In the rodent model
actuation due to ketamine-induced synchronous cortical activities coil, the maximum efficiency of −24.88 dB is located at the corners,
(Supplementary Fig. 12a). The amplitudes of the evoked responses decreasing to −28.90 dB at the centre. The power received at the
showed the well-known dependency (in electrophysiology litera- microimplant also depends on the horizontal and angular align-
ture44,45) on the stimulus pulse width, where longer pulses evoked ment of the Tx coils, as shown in Supplementary Fig. 13c,d. For
the largest LFP responses (Fig. 4g). In another neuromodulation the prospective primate model, the lateral movement (x–y plane)
example, we delivered fast 400 Hz stimulation pulse trains in 100 ms of the Tx coil results in a loss of 1.3 dB for a 3 mm horizontal dis-
blocks, and we were able to observe characteristic neuronal burst- placement. For the asymmetric coil morphology used in the in vivo
ing with increased post-stimulus firing rates described in the litera- rodent experiments, a y-directional misalignment of 3 mm resulted
ture43,46 under ketamine anaesthesia (demonstrated in Fig. 4h and in a 0.6 dB penalty, while an x-directional misalignment resulted in
Supplementary Fig. 12b). a much larger 6.73 dB loss, which is expected given the coil aspect
ratios. A 20° tilt resulted in 1.29 dB versus 1.90 dB loss in WPT
Characterizing wireless power link for optimal efficiency in the primate and rodent configurations, respectively.
performance In both simulations and measurements, a given neurograin was
To optimize the efficiency of electromagnetic data and power trans- in the centre of the subunit coil (Supplementary Fig. 13e,f shows
fer between the external hub and ensembles of spatially distributed the loss characteristics across the array). In the actual operation

Nature Electronics | www.nature.com/natureelectronics

Articles Nature Electronics

a b
Tx coil Tx coil
Cross-section Cross-section

Skin
8 mm Skin
5 mm
Skull Skull

Relay Relay
coil Grey matter coil Grey matter
5 mm 5 mm

c d Magnetic field of the relay coils

B (a.u.) B (a.u.)
1 1

10 mm
–1 –1

e f

Efficiency (dB)
–14

(0.8 mm unit)
Simulation Measurement 1

Simulation
–31

Y grid
8 –22 8 –16
2
–18
Y grid (1.0 mm unit)

Y grid (1.0 mm unit)

–32 3
–24
Efficiency (dB)

6 6 Efficiency (dB)
–33 2 4 6 8 10
–26 –34 X grid (1.0 mm unit)
4 4

Efficiency (dB)
Measurement –25

(0.8 mm unit)
–35 1
–28 Y grid –26
2 2 2 –27
–36
–30 3 –28

2 4 6 8 2 4 6 8 2 4 6 8 10
X grid (1.0 mm unit) X grid (1.0 mm unit) X grid (1.0 mm unit)

g h In phase Out of phase i 30

1.0 1.0
1
10
28

Clock frequency (MHz)

0.8 0.8
Amplitude (a.u.)

Amplitude (a.u.)

26
BER (%)

0.6 0.6
0
10 24
0.4 0.4

0.2 0.2 22

0 0 10–1 20
0 200 0 200 In Out of In Out of
Time (µs) Time (µs) phase phase phase phase

Fig. 5 | Wireless efficiency characterization. a, WPT system proposed for a primate model consists of the transmitting (Tx) coil, relay coil and receiver
(Rx) microcoils targeting 8 mm separation (4 mm skin and 4 mm skull). b, Coils for the rodent model targeting 5 mm separation (2.5 mm skin and 2.5 mm
skull). c, Photographs of the packaged Tx and relay coil for the primate and rodent model with an embedded chip array. d, Magnetic flux (B) generated
by the relay coil design that shows alternating fluxes of subunit coils (the direction of flux is shown as red or blue). e,f, Maps of simulated (left, top) and
measured (right, bottom) wireless transfer efficiency with regard to the spatial location inside one of the subunit relay coil for the primate and rodent
model, respectively. The X grid size is 1.0 mm for both cases and the Y grid size is 1.0 mm for the primate model coil and 0.8 mm for the rodent model coil.
The magnetic field distribution of one subcoil is shown for the symmetrical structure. g, Parallel-relay scheme with 69 chips activated through two adjacent
relay coils with 8 mm separation. In the current coil design, multiple relay coils placed in proximity to each other do not destructively interfere, allowing
scalability of the wireless power and data communication area. h, Received RF data packet shapes using shared RF carrier sources in phase (left) or
two separate RF carrier sources out of phase (right). The out-of-phase condition simulates the interference of two regular coils under non-ideal phase
shifting. i, BER (left) and clock frequency (right) of n = 865 and n = 733 data packets (in phase and out of phase, respectively). Bar charts show the average
values and the error bars are presented as mean values ± s.e.m.

of the neurograin system, we used the strength of the backscat- ple of such scalability, Fig. 5g demonstrates a dual Tx and dual Rx
tered RF signal as the guide to optimally align the Tx coil with channel, 69-node system using autonomous TDMA chips. When
submillimetre accuracy. the dual Tx coils are in phase, implemented simply through sharing
This wireless powering and data transmission scheme is intrinsi- the RF source in our WPT design, each subunit coil drives the cur-
cally designed to be parallel-extendable to cover multiple cortical rent in a direction opposite to that of the adjacent coils. Therefore,
areas and increase the channel capacity in the future. As an exam- their respective magnetic fluxes do not destructively co-interfere,

Nature Electronics | www.nature.com/natureelectronics

Nature Electronics Articles
as depicted by the retained packet shapes in Fig. 5h, left, and the showed that the multichannel system can be used in vivo in a rat
averaged BER is 0.5% over 865 packets. We simulated the situation model to record epicortical ECoG signals and intracortically stimu-
where the phase tuning is not ideal by driving two Tx coils with late neuronal microcircuits. While the head size of the rodent model
separate non-phase-locked RF carrier sources. When the two Tx has limited us here to 48 implanted neurograins, our network-
coils are out of phase (the typical case for two normal coils), the ing approach has the potential to scale up to 770 nodes. The elec-
destructive interference of their magnetic fields results in deformed tronic design of the neurograin SoC circuits can also, in principle,
packet shapes (Fig. 5h, right). The mean BER over 733 data pack- be ported to further deep submicrometre CMOS process nodes
ets then rises to 15.9% (collided packets are excluded for both (such as the 22 nm node), reducing the chip volume further. This
configurations). The average clock frequencies across the chips could provide a route to the minimally obtrusive implantation of
are similar in both conditions (Fig. 5i, right), which suggests that large populations of devices using recently developed intracorti-
our dual-channel system—only using a common RF source—can cal insertion techniques50. Importantly, our bidirectional wireless
deliver wireless energy to the larger number of spatially dispersed communication approach lends itself to real-time adaptive sensing:
nodes without the risk of notable cross-talk or BER deterioration that is, focusing on subpopulations of neurograins that are the most
(‘Dual-transmitter wireless power transfer’ in the Supplementary directly engaged in information exchange with the underlying bio-
Information provides a more detailed description). logical circuits.

Assessment of specific absorption rate of wireless link Methods

In considering the current regulatory limits for RF power, Equivalent circuit design for two- and three-coil wireless system. In a two-coil
Supplementary Fig. 14 shows the computed specific absorption rate system (shown in Supplementary Fig. 15a), the efficiency of wireless power delivery
(η) between the Tx coil and Rx coil can be described by
(SAR) maps averaged over 10 g of tissue47. According to the pres-
ent IEEE guidance (and the related FDA dictum), adverse effects ηtwo-coil =
k212 Q1 Q2L
, (1)
are not expected below the 10 W kg–1 SAR averaged over 10 g of the 1 + k212 Q1 Q2L
tissue ‘threshold’ value48. Applying this standard, our primate coil
system can transfer enough power for current ASICs to operate where k12 is the coupling factor between the Tx and Rx coil; further, Q1 = ω0L1/R1
and Q2L = Q2/(Q2QL + 1), where Q2 = ω0L2/R2 and QL = RL/ω0L2 (refs. 51,52); ω0 is
to 48.4% of the lateral relay coil area, which can include approxi-
angular frequency, L1 and R1 are the inductance and parasitic resistance of the Tx
mately 500 neurograins (Supplementary Fig. 14a). For the rodent coil, and L2 and R2 are the inductance and parasitic resistance of the Rx coil. Here,
coil and values simulated for our experiments, an input power of due to load transform from the resonant capacitor C2, RL can be solved as ω20 L22 /Rc,
0.072 W (18.6 dBm) would provide sufficient RF energy to the neu- where Rc is the circuit equivalent impedance34. Also, k12 depends on the size of
rograins across the relay coil, at the equivalent peak spatial SAR of the coils and the distance and arrangement between the two coils. In the case of
the implanted on-chip microcoils, for example, on the surface of the cortex, the
3.69 W kg–1 (Supplementary Fig. 14b). The actual measured WPT size mismatch between the Tx and Rx antennas is considerable (where the latter
efficiency was, on average, 7.4 and 9.8 dB lower than the simulated are ~0.25 mm2 in our system) so that given the separation between the two coils
values in the primate and rat model coil, respectively, although the exceeding several millimetres, the maximum k12 is severely limited.
spatial dependence of the measured efficiency generally followed In the three-coil system, a secondary coil can be introduced so that it can have
the simulated results. We note that the power measurement includes moderate coupling with both Tx and Rx coils and thus can increase the overall
efficiency of WPT (Supplementary Fig. 15b). The coupling factor between the relay
losses at multiple stages across the link from the RF source to the and Rx coil, kR2, can be ten times larger than k12 in the co-planar arrangement and
on-chip circuit. The overall efficiency can be further increased, the efficiency in three-coil wireless powering system can be simply written as
for example, by improving the on-chip rectifier efficiency, or bet-
(k21R Q1 QR )(k2R2 QR Q2L )
ter tailoring the impedance match in the coil/RF circuitry. Due to ηthree-coil =
(1+k21R Q1 QR +k2R2 QR Q2L )(1+k2R2 QR Q2L )
the specifics of the TSMC 65 nm foundry process, the embedded [ ] [ ]
planarization-enabling filling structures across the multiple metal =
( k21R Q1 QR
)
×
(k2R2 QR Q2L ) (2)
( 1+k21R Q1 QR +k2R2 QR Q2L ) ( 1+k2R2 QR Q2L )
layers of the ASIC chip may have further limited the cross-section
and coupling factor of the Rx coil49. Further improvement in the = η1R × ηR2 ,
peak SAR will be dependent on identifying the sources of the
where k1R is the coupling factor between the Tx and relay coil51 and QR is the
reported additional 7 dB measured loss in our system and mitigat- unloaded quality factor of the relay coil. Using a secondary coil that has a size
ing these causes. When achieving the simulated maximal efficiency, comparable to the Tx coil, k1R can be increased, and a high η1R can be achieved; η1R
0.1 W (20 dBm) of Tx power in the primate model coil can pro- is the WPT efficiency between the Tx and relay coil. Also, co-planar placement of
vide adequate energy to a neurograin ASIC at any location within the relay and Rx coils can improve kR2, which results in higher ηR2; ηR2 is the WPT
the relay coil area, which corresponds to a peak spatial SAR of efficiency between the relay and Rx coil. Note that since the relay coil is not directly
loaded by the source or load impedance, η1R and ηR2 benefits from a high quality
3.66 W kg–1, averaged over 10 g according to results. factor for the relay coil QR. Consequently, the relay coil collects high magnetic
Lastly, Supplementary Fig. 14c shows the effect of an extra adja- energy to the target area, improving the power transfer efficiency for the ensemble
cent Tx/relay coil on the SAR value, demonstrating that adding submillimetre Rx coils. Based on the simulated coil parameters in Supplementary
another RF WPT/data path (for example, for wider cortical cover- Table 2 and equations (1) and (2), the maximum and minimum efficiencies
age across multiple areas) does not affect the local SAR value. Thus, within the relay coil are calculated in the primate and rodent model for two- and
three-coil systems, as shown in Supplementary Fig. 15c. The calculated efficiency
from the SAR point of view, our wireless system should be able to complies with the three-coil simulation (Fig. 5e,f) and also reveals more than 15 dB
scale up using multiple RF paths (that is, scaling for neurograin efficiency increase in the three-coil system compared with the two-coil system.
populations in the thousands) with a minimal impact on the peak
local SAR values while distributing the RF power. Electromagnetic simulation. The design of the coils was performed and
simulated on a high-frequency structure simulator (HFSS Ansys electromagnetic
suite, Ansys) with tissue models including their dielectric properties around
Conclusions 1 GHz. The design parameters for the Tx and relay coil were chosen to maximize
We have reported a scalable, implantable wireless microsensor and the WPT efficiency within the power projection area, while the maximum
microstimulator network. Each submillimetre-sized silicon chip area is limited by the self-resonance of the coils. In the rodent WPT system, the
(0.1 mm3)—or neurograin—combines RF energy harvesting, data size of the coil was limited by the size of the rat skull. The on-chip Rx coil was
communication, and recording or stimulation capabilities in an also designed via electromagnetic simulation to improve the WPT efficiency and
impedance matching with the high-load/low-power ASIC. The S-parameters
autonomous unit with a unique ID. We designed a time-division from the HFSS simulation results were imported into the Mixed-Signal Simulator
multiplexing approach to build a spatially distributed microsen- (Cadence Design Systems) to set up a full RF simulation test bench for wireless
sor system that communicates with an external wireless hub. We circuit design.

Nature Electronics | www.nature.com/natureelectronics

Articles Nature Electronics
Post-processing and encapsulation of microimplants. Sixty-micrometre pads The insertion of neurograin electrodes was manually done, where the neurograin
on a bare ASIC offer an electrical path to biological tissue, while their aluminium array and relay coil were fixed onto the skull with dental cement, if needed. The
surface is not suitable for long-term tissue interface. For recording neurograins, Tx coil and coaxial cable connecting to the RF hub were held at the target position
a 60 μm × 60 μm gold electrode surface can provide an appropriate impedance by a clamp stand. For interoperative recording and stimulation measurements,
(300 kΩ at 1 kHz) to meet the input requirement of capacitor-less AFE for isoflurane was switched to 80 mg kg–1 ketamine and 5 mg kg–1 xylazine mixture, and
ECoG recording. Further, tungsten penetrating electrodes have a sufficiently the depth of anaesthesia was monitored every 5 min by checking the breathing rate
low impedance for neurograin stimulation—a nominal 10 kΩ impedance at and paw retraction responses. If needed, one-fourth of the initial dosage was given
1 kHz. Using a polymer-based customized technique, we embedded chips into to the animal to maintain a stable anaesthesia level. The duration of recording and
PDMS, which allows photolithography38. After photoresist patterning, we directly stimulation experiment under the ketamine/xylazine mixture was approximately
deposited 50 nm titanium and 200 nm gold on the pad to serve as the electrodes, 2 h. A heating pad was placed below the animal to maintain the body temperature,
where the electrodes are defined by the photoresist during the lift-off process. and saline was occasionally administrated to the cortex to prevent the drying
For the stimulation chip, 1.5-mm-long tungsten wires with 50 μm diameter were of tissue. On the completion of surgery, the animals were euthanized with
manually attached to the pads with silver epoxy (EPO-TEK H20E) and covered pentobarbital sodium.
with an epoxy resin (Scotch-Weld DP100). After electrode fabrication, the implants
were encapsulated by parylene C, except for electrical contacts for recording or Reporting Summary. Further information on research design is available in the
stimulation to achieve long-term biocompatibility. Neurograins and the assembly Nature Research Reporting Summary linked to this article.
were sterilized with ultraviolet radiation or ethylene oxide.
Data availability
Coil fabrication, assembly and packaging. The Tx and relay coils were fabricated The dataset that supports the plots within this paper and other findings of this
on a polyimide substrate with a 0.1 mm thickness under a two-layer PCB process study are available in the Supplementary Information and from the corresponding
using a two ounce copper trace. For the Tx coil, a trace width of up to 1.0 mm was author upon reasonable request.
used to improve the quality factor, while in the relay coil, a trace width of 0.2 mm
is used to reduce the effect of the surrounding dielectric tissue. These design
parameters were optimized in electromagnetic simulation before the fabrication. Code availability
The coils were then assembled with matching capacitors and encapsulated with Custom-developed MATLAB codes for demodulation are available in the
PDMS (Dow SYLGARD 182) or 3M biomedical-grade epoxy resin. For the Tx Supplementary Information and from the corresponding author upon reasonable
coils, 50 Ω mini-RF coaxial connector and cable are also attached. PDMS package request.
processing was done with an acrylic mould. Before the packaging, coils were cleaned
with isopropyl alcohol, acetone and deionized water, and the mould was coated Received: 14 September 2020; Accepted: 9 July 2021;
with an antiadhesion spray. The base PDMS layer with 0.5 mm thickness was cured Published: xx xx xxxx
for 3 h at 100 °C, and attached to polyimide coils. The top PDMS layer was covered
again, resulting in an overall packaging thickness of 1 mm. The microimplant
chips are tested without the package or packaged in PDMS or parylene C. When
References
1. Hochberg, L. R. et al. Neuronal ensemble control of prosthetic devices by a
longer-term hermetic packaging is needed, the relay coils are encapsulated with a
human with tetraplegia. Nature 442, 164–171 (2006).
liquid crystal polymer by using a metal jig and a press machine53.
2. Afshar, A. et al. Single-trial neural correlates of arm movement preparation.
Neuron 71, 555–564 (2011).
Data telemetry and demodulation. An SDR (Raptor, Rincon Research or 3. Truccolo, W., Friehs, G. M., Donoghue, J. P. & Hochberg, L. R. Primary
FMCOMMS3, Analog Devices) consisting of an Analog Devices AD9361 RF motor cortex tuning to intended movement kinematics in humans with
transceiver and a Zynq SoC was used to generate the RF transmission energy tetraplegia. J. Neurosci. 28, 1163–1178 (2008).
and to communicate with the ASICs54. The SDR generates a transmitting tone 4. Nurmikko, A. V. et al. Listening to brain microcircuits for interfacing with
at 915 MHz, which is further amplified with an integrated-circuit-based power external world—progress in wireless implantable microelectronic
amplifier (ADL5605-EVALZ, Analog Devices) board. An RF surface acoustic wave neuroengineering devices. Proc. IEEE 98, 375–388 (2010).
duplexer (D5DA942M5K2S2, Taiyo Yuden) isolates the transmitting and receiving 5. Rubehn, B., Bosman, C., Oostenveld, R., Fries, P. & Stieglitz, T. A
paths. The SDR receiver front end performs signal amplification, downconversion MEMS-based flexible multichannel ECoG-electrode array. J. Neural Eng. 6,
(945 MHz) and digitization at 30–45 MSa s–1. The digitized IQ data were then 036003 (2009).
ported to a personal computer for offline processing (in MATLAB/Simulink), 6. Pistohl, T., Ball, T., Schulze-Bonhage, A., Aertsen, A. & Mehring, C.
which can be embedded into the Zynq SoC for onboard demodulation. The Prediction of arm movement trajectories from ECoG-recordings in humans.
digital bitstream is decoded from the digitized IQ data after filtering, automatic J. Neurosci. Methods 167, 105–114 (2008).
gain control, coarse and fine frequency compensation, bit strobe correction and 7. Jun, J. J. et al. Fully integrated silicon probes for high-density recording of
BPSK demodulation. These processes have been optimized for the neurograin neural activity. Nature 551, 232–236 (2017).
communication scheme. Automatic gain control normalizes the uplink data 8. Musk, E. et al. An integrated brain-machine interface platform with
packet power, which varies between chips. Frequency compensation and bit strobe thousands of channels. J. Med. Internet Res. 21, e16194 (2019).
correction processes, which recover the clock frequency of each chip, compensate 9. Miranda, H., Gilja, V., Chestek, C. A., Shenoy, K. V. & Meng, T. H.
the clock variance between chips to improve the demodulation performance. HermesD: a high-rate long-range wireless transmission system for
simultaneous multichannel neural recording applications. IEEE Trans.
RF wireless experiment setup. A liquid head phantom with a permittivity of 41.8 Biomed. Circuits Syst. 4, 181–191 (2010).
and conductivity of 0.97 S m–1 around 900 MHz was made in accordance with 10. Borton, D. A., Yin, M., Aceros, J. & Nurmikko, A. An implantable
the IEEE Standard 1528-2013 to mimic the dielectric properties of head tissue47. wireless neural interface for recording cortical circuit dynamics in moving
The liquid phantom consisted of 64.81% 1,2-propanediol, 0.79% NaCl and 34.4% primates. J. Neural Eng. 10, 026010 (2013).
deionized water. A phantom bath was designed for the wireless test using acrylic 11. Yin, M., Borton, D. A., Aceros, J., Patterson, W. R. & Nurmikko, A. V.
plates and a 0.1-mm-thick plastic sheet to reduce RF interference (Supplementary A 100-channel hermetically sealed implantable device for chronic
Fig. 16). Separation between the Tx and relay coils was 8 mm or 5 mm and the wireless neurosensing applications. IEEE Trans. Biomed. Circuits Syst. 7,
gap was filled with the liquid phantom. Two plastic sheets supported the liquid 115–128 (2013).
phantom and relay coil. For the efficiency measurement experiment, grid paper 12. Simeral, J. D. et al. Home use of a wireless intracortical brain-computer
was attached inside the relay coil to guide the manual placement of the neurograin interface by individuals with tetraplegia. IEEE Trans. Biomed. Eng. 68,
chip. For the Tx alignment, a three-dimensionally printed Tx coil holder was 2313–2325 (2021).
placed below the bath and above the paper with grid marks, which provides 13. Song, Y.-K. et al. Active microelectronic neurosensor arrays for implantable
references for alignment between the Tx and relay coil. brain communication interfaces. IEEE Trans. Neural Syst. Rehabil. Eng. 17,
339–345 (2009).
In vivo experiment. Two wild-type Long–Evans adult male rats (500–700 g) 14. Zanos, S., Richardson, A. G., Shupe, L., Miles, F. P. & Fetz, E. E. The
were used in this study for neurograin recording and stimulation experiments. Neurochip-2: an autonomous head-fixed computer for recording and
All the research protocols were approved and monitored by the Brown University stimulating in freely behaving monkeys. IEEE Trans. Neural Syst. Rehabil.
Institutional Animal Care and Use Committee, and all research was performed in Eng. 19, 427–435 (2011).
accordance with relevant NIH guidelines and regulations. Rats were anaesthetized 15. Liu, X. et al. The PennBMBI: design of a general purpose wireless
with 2–3% isoflurane throughout the surgery. After shaving the hair, the animal brain-machine-brain interface system. IEEE Trans. Biomed. Circuits Syst. 9,
was fixed on a stereotaxic frame, and the head skin was sterilized. Up to 5 cm 248–258 (2015).
sagittal incision was made in the skin over the skull, and a bilateral craniotomy was 16. Zhou, A. et al. A wireless and artefact-free 128-channel neuromodulation
performed using a precision surgery dental drill. After the craniotomy, the skull in device for closed-loop stimulation and recording in non-human primates.
the midline was thinned down to improve contact with the neurograin interface. Nat. Biomed. Eng. 3, 15–26 (2019).

Nature Electronics | www.nature.com/natureelectronics

Nature Electronics Articles
17. Seo, D., Carmena, J. M., Rabaey, J. M., Maharbiz, M. M. & Alon, E. Model 44. Brindley, G. S. & Lewin, W. S. The sensations produced by electrical
validation of untethered, ultrasonic neural dust motes for cortical recording. stimulation of the visual cortex. J. Physiol. 196, 479–493 (1968).
J. Neurosci. Methods 244, 114–122 (2015). 45. Kim, S. et al. Behavioral assessment of sensitivity to intracortical
18. Seo, D. et al. Wireless recording in the peripheral nervous system with microstimulation of primate somatosensory cortex. Proc. Natl Acad. Sci. USA
ultrasonic neural dust. Neuron 91, 529–539 (2016). 112, 15202–15207 (2015).
19. Piech, D. K. et al. A wireless millimetre-scale implantable neural stimulator 46. Fi´ath, R. et al. Large-scale recording of thalamocortical circuits: in vivo
with ultrasonically powered bidirectional communication. Nat. Biomed. Eng. electrophysiology with the two-dimensional electronic depth control silicon
4, 207–222 (2020). probe. J. Neurophysiol. 116, 2312–2330 (2016).
20. Jia, Y. et al. A mm-sized free-floating wirelessly powered implantable optical 47. Determining the peak spatial-average specific absorption rate (SAR) in the
stimulation device. IEEE Trans. Biomed. Circuits Syst. 13, 608–618 (2019). human body from wireless communications devices, 30 MHz to 6 GHz—part
21. Jia, Y. et al. A dual-band wireless power transmission system for evaluating 3: specific requirements for using the finite difference time domain (FDTD)
mm-sized implants. IEEE Trans. Biomed. Circuits Syst. 13, 595–607 (2019). method for SAR calculations of mobile phones. IEC/IEEE 62704-3:2017
22. Gutruf, P. et al. Fully implantable optoelectronic systems for battery-free, 1–76 (2017).
multimodal operation in neuroscience research. Nat. Electron. 1, 48. IEEE standard for safety levels with respect to human exposure to radio
652–660 (2018). frequency electromagnetic fields, 3 kHz to 300 GHz. IEEE Std C95.1-2005
23. Ho, J. S. et al. Wireless power transfer to deep-tissue microimplants. Proc. (Revision of IEEE Std C95.1-1991) 1–238 (2006).
Natl Acad. Sci. USA 111, 7974–7979 (2014). 49. Shilimkar, V. S., Gaskill, S. G. & Weisshaar, A. Experimental characterization
24. Khalifa, A. et al. The microbead: a 0.009 mm3 implantable wireless neural of metal fill placement and size impact on spiral inductors. In 2009 IEEE 18th
stimulator. IEEE Trans. Biomed. Circuits Syst. 13, 971–985 (2019). Conference on Electrical Performance of Electronic Packaging and Systems
25. Yeon, P., Bakir, M. S. & Ghovanloo, M. Towards a 1.1 mm2 free-floating 101–104 (IEEE, 2009).
wireless implantable neural recording SoC. In 2018 IEEE Custom Integrated 50. Sigurdsson, S. A., Yu, Z., Lee, J. & Nurmikko, A. A method for large-scale
Circuits Conference (CICC) 1–4 (IEEE, 2018). implantation of 3D microdevice ensembles into brain and soft tissue.
26. Agrawal, D. R. et al. Conformal phased surfaces for wireless powering of Microsyst. Nanoeng. 6, 97 (2020).
bioelectronic microdevices. Nat. Biomed. Eng. 1, 0043 (2017). 51. Kiani, M., Jow, U.-M. & Ghovanloo, M. Design and optimization of a 3-coil
27. Lee, J. et al. An implantable wireless network of distributed microscale inductive link for efficient wireless power transmission. IEEE Trans. Biomed.
sensors for neural applications. In 2019 9th International IEEE/EMBS Circuits Syst. 5, 579–591 (2011).
Conference on Neural Engineering (NER) 871–874 (IEEE, 2019). 52. Finkenzeller, K. RFID Handbook: Fundamentals and Applications in
28. Lee, J. et al. Wireless power and data link for ensembles of sub-mm scale Contactless Smart Cards, Radio Frequency Identification and Near-Field
implantable sensors near 1Ghz. In 2018 IEEE Biomedical Circuits and Systems Communication (John Wiley & Sons, 2010).
Conference (BioCAS) 1–4 (IEEE, 2018). 53. Jeong, J. et al. A miniaturized, eye-conformable, and long-term reliable
29. Leung, V. W. et al. Distributed microscale brain implants with wireless power retinal prosthesis using monolithic fabrication of liquid crystal polymer
transfer and Mbps bi-directional networked communications. In 2019 IEEE (LCP). IEEE Trans. Biomed. Eng. 62, 982–989 (2014).
Custom Integrated Circuits Conference (CICC) 1–4 (IEEE, 2019). 54. Cai, H. et al. A software-defined radio for wireless brain implants network.
30. Laiwalla, F. et al. A distributed wireless network of implantable sub-mm In Proc. 24th Annual International Conference on Mobile Computing and
cortical microstimulators for brain-computer interfaces. In 2019 41st Annual Networking 852–854 (Association for Computing Machinery, 2018).
International Conference of the IEEE Engineering in Medicine and Biology
Society (EMBC) 6876–6879 (IEEE, 2019).
31. Ha, S. et al. Silicon-integrated high-density electrocortical interfaces. Proc. Acknowledgements
IEEE 105, 11–33 (2016). We thank Y.-K. Song for insights into microfabrication and ASIC design, J. Jeong for
32. Ahmadi, N. et al. Towards a distributed, chronically-implantable neural his expertise with neurograin hermetic sealing and materials processing, and R. Rao
interface. In 2019 9th International IEEE/EMBS Conference on Neural for guidance on wireless networking design. We also thank C. Kilfoyle, E. Mok, S.
Engineering (NER) 719–724 (IEEE, 2019). Sigurdsson, and the Animal Care facility at Brown University for their contributions. We
33. Spilker, J. J. Jr. Digital Communications by Satellite (Prentice-Hall, 1977). also acknowledge S. Li, S. Yu, L. Cui, S. Alluri and M. Lokhandwala at UCSD for their
34. Ahn, D. & Ghovanloo, M. Optimal design of wireless power transmission work on ASIC design and benchtop test. We have greatly benefited from the insight of
links for millimeter-sized biomedical implants. IEEE Trans. Biomed. Circuits our colleagues across multiple fields from microelectronics to brain sciences and clinical
Syst. 10, 125–137 (2015). neurology: J. Donoghue, B. Dutta, J. Groe, L. Hochberg, T. Sejnowsky, K. Shenoy and
35. Feng, P., Maslik, M. & Constandinou, T. G. EM-lens enhanced power S. Cash, among others. This research was initially supported by Defense Advanced
transfer and multi-node data transmission for implantable medical devices. Research Projects Agency N66001-17-C-4013, with subsequent support from private
In 2019 IEEE Biomedical Circuits and Systems Conference (BioCAS) gifts. The post-processing work benefited from the equipment funded by MRI award no.
1–4 (IEEE, 2019). DMR-1827453.
36. Theilmann, P. T., Presti, C. D., Kelly, D. J. & Asbeck, P. M. A µW
complementary bridge rectifier with near zero turn-on voltage in SOS CMOS Author contributions
for wireless power supplies. IEEE Trans. Circuits Syst. I, Reg. Papers 59, A.N. conceived the project with P.A., P.P.M. and L.L. J.L. and A.N. designed the
2111–2124 (2012). neural experiment concept. J.L. designed the RF wireless powering and data
37. Leung, V. W. et al. A CMOS distributed sensor system for high-density communication method, performed system characterization and in vivo experiments,
wireless neural implants for brain-machine interfaces. In ESSCIRC and analysed the results. V.L., P.A., S.S. and L.L. designed the RF data communication
2018—IEEE 44th European Solid State Circuits Conference (ESSCIRC) protocol. V.L., J.H., F.L. and P.P.M. designed the neurograin ASICs. A.-H.L. performed
230–233 (IEEE, 2018). microfabrication and packaging. J.L., V.L., F.L. and A.N. wrote the paper. All the authors
38. Lee, A.-H. et al. A scalable and low stress post-CMOS processing technique commented on the paper.
for implantable microsensors. Micromachines 11, 925 (2020).
39. Yang, K., Dong, Q., Blaauw, D. & Sylvester, D. 8.3 A 553F2 2-transistor
amplifier-based physically unclonable function (PUF) with 1.67% native Competing interests
instability. In 2017 IEEE International Solid-State Circuits Conference (ISSCC) The authors declare no competing interests.
146–147 (IEEE, 2017).
40. Suh, G. E. & Devadas, S. Physical unclonable functions for device Additional information
authentication and secret key generation. In 2007 44th ACM/IEEE Design Supplementary information The online version contains supplementary material
Automation Conference 9–14 (IEEE, 2007). available at https://doi.org/10.1038/s41928-021-00631-8.
41. Huang, J. et al. A 0.01-mm2 mostly digital capacitor-less AFE for distributed
autonomous neural sensor nodes. IEEE Solid-State Circuits Lett. 1, Correspondence and requests for materials should be addressed to A.N.
162–165 (2018). Peer review information Nature Electronics thanks Nitish Thakor, Samantha Santacruz
42. Jeong, J. et al. Conformal hermetic sealing of wireless microelectronic and Jonathan Viventi for their contribution to the peer review of this work.
implantable chiplets by multilayered atomic layer deposition (ALD). Adv. Reprints and permissions information is available at www.nature.com/reprints.
Funct. Mater. 29, 1806440 (2019).
43. Steriade, M., Contreras, D., Amzica, F. & Timofeev, I. Synchronization of fast Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
(30–40 Hz) spontaneous oscillations in intrathalamic and thalamocortical published maps and institutional affiliations.
networks. J. Neurosci. 16, 2788–2808 (1996). © The Author(s), under exclusive licence to Springer Nature Limited 2021

Nature Electronics | www.nature.com/natureelectronics

 nature research | reporting summary
Corresponding author(s): Arto Nurmikko
Last updated by author(s): Jul 4, 2021

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection For this study the following software was used to collect data and program devices: MATLAB (version R2019b), C (version C17), Arduino IDE
(version 1.8), HFSS (Ansys, version 2020)

Data analysis Data analysis and plotting was executed with MATLAB (version R2019b).
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
April 2020

- A description of any restrictions on data availability

Codes and the dataset that support the plots within this paper and other findings of this study are available in the supplementary file and from the corresponding
author upon reasonable request.

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size Sample size was chosen considering the major focus of this study is in the hardware validation and the collected data was compared with that
from the commercial hardware and other literatures.

Data exclusions No data was excluded from the analysis

Replication All experiments has been performed in tissue-proxy (saline) condition or the animal independently and replication of system interface was
successful. Also, multiple data points have been collected in the case of biology studies.

Randomization This study focuses on the hardware system validation and no randomization study was needed.

Blinding This study focuses on the hardware system validation and no binding study was needed.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology and archaeology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data
Dual use research of concern

Animals and other organisms

Policy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research
Laboratory animals Wild-type Long-Evans adult male rats (500–700 g) with the age of 8-12 weeks were used in this study

Wild animals No wild animals were used in the study

Field-collected samples No field collected samples were used in the study

Ethics oversight All research protocols were approved and monitored by the Brown University Institutional Animal Care and Use Committee

Note that full information on the approval of the study protocol must also be provided in the manuscript.
April 2020