S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

PHYSICS Copyright © 2025 The

Authors, some rights
Tensor networks enable the calculation of turbulence reserved; exclusive
licensee American
probability distributions Association for the
Advancement of
Science. No claim to
Nikita Gourianov1,2*, Peyman Givi1, Dieter Jaksch2,3, Stephen B. Pope4 original U.S.
Government Works.
Predicting the dynamics of turbulent fluids has been an elusive goal for centuries. Even with modern computers, Distributed under a
anything beyond the simplest turbulent flows is too chaotic and multiscaled to be directly simulatable. An alterna- Creative Commons
tive is to treat turbulence probabilistically, viewing flow properties as random variables distributed according to Attribution License 4.0
joint probability density functions (PDFs). Such PDFs are neither chaotic nor multiscale, yet remain challenging to (CC BY).
simulate due to their high dimensionality. Here, we overcome the dimensionality problem by encoding turbu-
lence PDFs as highly compressed “tensor networks” (TNs). This enables single CPU core simulations that would
otherwise be impractical even with supercomputers: for a 5 + 1 dimensional PDF of a chemically reactive
( ) turbu-
lent flow, we achieve reductions in memory and computational costs by factors of  106 and  103 , respec-
( )

tively, compared to standard finite-­difference algorithms. A future path is opened toward something heretofore
thought infeasible: directly simulating high-­dimensional PDFs of both turbulent flows and other chaotic systems
that can usefully be described probabilistically.

Downloaded from https://www.science.org on February 07, 2025

INTRODUCTION flows (6, 7), direct schemes like finite differences (FDs) or volumes
Despite the simple and deterministic physical laws governing it, tur- were long ago dismissed as computationally infeasible (8) due to
bulence remains an inherently complex and chaotic phenomenon. It is their seemingly exponential cost in d. This spurred the creation of
characterized by large numbers of eddies interacting in intricate and indirect Monte Carlo (MC) algorithms for probabilistic turbulence
nonlinear ways across wide ranges of spatial and temporal scales, lead- simulations (8). These schemes have proven highly successful, enabling
ing to the emergence of chaos. Making matters worse, practically advanced turbulent combustion simulations involving thousands of
important turbulent flows (e.g., fuel-­oxidizer mixtures in combustion) CPU cores (9, 10). However, the randomness and slow convergence
often involve multiple chemically reacting species, which introduces characterizing MC methods can be avoided by directly solving the
additional nonlinearities and scales. The presence of chaos prohibits underlying Fokker-­Planck equations.
predicting the exact dynamics of turbulent flow fields over long peri- It is not just probabilistic turbulence calculations that are hin-
ods of time, while the multiscaled nature of the flow fields makes their dered by the curse of dimensionality: quantum many-­body systems
simulation immensely expensive due to the need to solve sets of cou- are described by states whose sizes also grow exponentially (in the
pled partial differential equations (PDEs) on very fine grids. number of particles). However, physically relevant quantum states
However, for practical applications it is rarely necessary to know are known to be highly structured (11). Such structure can be ex-
the precise state of a turbulent flow field at every point in space-­time. ploited to compress the states into approximate, but highly accurate,
Rather, one is typically more interested in far slower-­varying statisti- polynomially large representations known as tensor networks (TNs).
cal quantities where the fluctuations are averaged out (such as the lift TN algorithms allow efficiently evolving these states and analyzing
and drag of an aeroplane or the rate of product formation in a chem- their physical properties without ever leaving the compressed TN
ical process). In the statistical description of turbulence, variables representation (12–15) and have enabled the simulation of otherwise
like velocities U, chemical mass-­fractions Φα, temperatures, etc., are intractable quantum systems like superconductors, ferromagnets,
treated as random variables (RVs) distributed according to some and quantum computers (16–24). Recently, the TN formalism has
one-­point, one-­time joint probability density function (PDF) (1) begun spreading beyond quantum physics (25–31).
( ) Decades of empirical experience indicates that f is also highly
f = f u, φ1 ,… ; x, t (1) structured. For instance, in homogeneous turbulence, velocities U are
across space x and time t, with u, φ1 being sample-­space variables often distributed normally (2), whereas mass fractions Φα have been
corresponding to U, Φ1. The trajectory of f completely describes the observed to follow normal, exponential, and β distributions in nonre-
one-­point, one-­time statistics of the flow dynamics (2), which are the active flows (32). In more complicated reacting flows, the PDFs gen-
central quantities of interest in practical engineering calculations. erally cannot be so simply parameterized (33), although they remain
The time evolution of f is modeled by Fokker-­Planck PDEs that smoother and more predictable than the underlying flow fields (34).
are straightforward to derive (3–5) but hard to solve. If f is d-­ This work shows that the structure contained in turbulence
dimensional, assigning M points for each dimension(results ) in a total PDFs is readily exploitable through TNs: using a simple TN known
of M d gridpoints. Given that d can be as high as  103 in realistic as the “matrix product state” (MPS) ansatz to encode f in a highly
compressed format allows us to formulate a scheme for cheaply
1
Department of Mechanical Engineering and Materials Science, University of and directly solving the governing Fokker-­Planck equations.
Pittsburgh, Pittsburgh, PA 15261, USA. 2Clarendon Laboratory, University of Oxford, When the PDF structure is well-­matched to the MPS ansatz, the
Oxford OX13PU, UK. 3Institut für Quantenphysik, Universität Hamburg, Luruper time-­e volution costs just ∼dlogM ; while standard FD schemes
Chaussee 149, 22761 Hamburg, Germany. 4Sibley School of Mechanical and Aerospace
Engineering, Cornell University, Ithaca, NY 14853, USA. scale as ∼M d . We demonstrate the advantage by looking at the
*Corresponding author. Email: nikgourianov@​icloud.​com following turbulent flow.

Gourianov et al., Sci. Adv. 11, eads5990 (2025) 29 January 2025 1 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

RESULTS in x along with the scalar-­ratio ⟨Φ1⟩ ∕ ⟨Φ2⟩ at four different times,
Probabilistic modeling of reactive turbulence showing how the initially orderly, unmixed flow state is driven
Consider an incompressible, three-­dimensional (3D) turbulent flow in toward a fully-­mixed ⟨Φ1⟩ ∕ ⟨Φ2⟩ ≈ 1 state by SGS and large-­scale
which two chemical species are irreversibly reacting: A + B → Products. convective and diffusive mixing. The SGS mixing leads to the PDF
In this system, two chemical species (of mass fractions Φ1 , Φ2) are concentrating, while the diffusion and mean-­flow convection in-
stirred by a velocity field U = U(x, t) in 3D space x across time t. For duces multimodality in the PDF. The MPS simulation is highly ac-
the sake of simplicity, we here consider only the statistics of the Φ1 , Φ2 curate (Fig. 2B), yet the
( number) of variables parameterizing the
scalar fields by assuming that the large-­scale statistical features of the PDF (NVPP) is only  1∕105 of an equivalent, classically imple-
hydrodynamics are known a priori, while modeling the subgrid-­scale mented FD scheme [Materials and Methods, Eq. 12].
(SGS), turbulent velocity fluctuations using large eddy simulation
(LES), per current best practices (10). Doing so eliminates the ran- MPS encoding
In our MPS) encoding, the discretized, high-dimensional f φ1 ,
(
domness( of U and ) reduces the dimensionality to d = 5 + 1. Now,
f = f φ1 , φ2 ; x, t describes the statistics of the mass fraction fluctua- φ2 , x1 , x2 , x3 is decomposed into a 1D chain of tensors, where the
tions, which provide the mean mass fractions ⟨Φ1⟩, ⟨Φ2⟩ through φ1 , φ2 , x1 , x2 , x3 dimensions are sequentially mapped to tensors from
left to right, with each dimension itself decomposed into multiple
tensors lengthscale by lengthscale {analogously to the “sequential,
� �
⟨Φα ⟩(x, t) =
∫[0,1]×2
φα f φ1 , φ2 ; x, t dφ1 dφ2 (2)
serial” ordering in [(27), Fig. 1] and [(38), Eq. 9]}. This encoding
Such PDFs are known as “filtered density functions” (3, 35). exploits two separate structures that characterize the solution of Eq.
Deriving the equation governing f requires SGS closure modeling. 3: First, the general smoothness ( of turbulence PDFs; second, that
the different dimensions of f φ1 , φ2 ; x, t are unlikely to be strongly
)
Using popular closure models (35) gives the Fokker-­Planck PDE

Downloaded from https://www.science.org on February 07, 2025

coupled at low CΩ, because for CΩ = 0 the PDF is separable (Supple-
mentary Text, “Separability of Fokker-­Planck equation” section).
� �
𝜕f 𝜕f 𝜕 � � 𝜕f 𝜕 � � � � 𝜕 � �
+ ⟨Ui ⟩ − γ + γSGS = Ωmix φα − ⟨Φα ⟩ f − Sα f (3)
𝜕t 𝜕 xi 𝜕 xi 𝜕 xi 𝜕 φα 𝜕 φα Matching the structure of the PDF in this way allows for an MPS
encoding that is both accurate and parsimonious. The MPS represen-
Here, U(x, t) is the large-­scale (or, “filtered”) mean hydrodynamic tation (like any TN) can be systematically compressed, i.e., the NVPP
[ ]×3
field across x ∈ 0, l0 , t ∈ 0, 2T0 , which is set to be a jet flow
[ ] reduced, by varying a hyperparameter known as the maximum bond
combined with a Taylor-­Green vortex of amplitude u0 = l0 ∕ T0 dimension χ. This hyperparameter regulates the maximal size of the
(Materials and Methods, “Flow case definition” section). “bonds” between the tensors, which is equivalent to the maximum
The left hand side of Eq. 3 denotes the PDF transport in space allowed coupling between the tensors and, in turn, between the differ-
and time. The first term is the rate of temporal change, and the ent lengthscales and dimensions of f . For example, setting χ = 1 for-
second term represents convection by the mean velocity field. bids any coupling between the tensors and makes the NVPP minimal,
The third represents the influence of the molecular (γ) and SGS while picking χ sufficiently large makes the MPS representation exact
diffusion [γSGS (x, t)] coefficients: The former sets the Peclet num- and NVPP = M 5 like in the standard representation. Setting χ to be
ber Pe = u0 l0 ∕ γ and the latter is modeled via the Smagorinsky small in turn leads to a low NVPP, but the MPS encoding will still re-
� � �2 main accurate if it reflects the structure of f sufficiently well.
Δ2 ∑ 𝜕Ui 𝜕Uj
(36) closure γSGS = Cs 2 + x
𝓁
ij x 𝜕 j
, with Cs an empirical
𝜕 i Validation of algorithm
We now investigate how well the MPS parameterization fits the solu-
constant and Δ 𝓁 the LES filter width (both are specified in Materials tion of Eq. 3 in practice. To determine the χ required to accurately
and Methods, “Flow case definition” section). simulate the dynamics of the RVs Φ1 , Φ2 underlying the PDF, the com-
The right hand side of Eq. 3 designates transport in the composi- position space transport parameters CΩ , Da are varied, while fixing the
tion space [“composition” since the φ1 , φ2 ∈ [0, 1]×2 mass fractions hydrodynamic variables ⟨ U ⟩, γSGS, and Pe to those used in Fig. 1.
define the composition of the fluid]. The first term represents scalar Increasing CΩ is expected to lead to higher coupling between the
mixing from the SGS turbulence and is modeled via the popular different dimensions of f , which reduces the efficiency of our MPS
γ+γ
least mean square estimation (LMSE) (37) closure Ωmix = CΩ Δ2SGS , encoding, i.e., increasing CΩ requires an increased χ to maintain ac-
𝓁 curacy. To verify, we first set Da = 0 because this allows us to accu-
with CΩ the SGS mixing rate. The final term denotes the effects of
rately compute ⟨Φα⟩ independently of Eq. 3 (see Materials and
chemical reaction. For the binary reaction scheme considered here,
Methods, “Moment equations” section) and benchmark the accu-
S1 = S2 = −Cr φ1 φ2, where Cr denotes the reaction rate that defines racy of the computed MPS-­PDF across CΩ , χ. The benchmark is
the Damköhler number Da = Cr l0 ∕ u0. shown in Fig. 2 (A and B). The ⟨Φ1⟩ ∕ ⟨Φ2⟩ ratios in Fig. 2A depict
To solve Eq. 3, we discretize f at every point in time on a M =128, d =5 how the MPS-­PDF means approach their numerically exact equiva-
Cartesian grid, but parameterize it as an MPS-­network using far lent when χ increases and CΩ decreases. All the cases, including the
fewer variables than the 1285 gridpoints resolving it. This allows us ones with lowest accuracy, correctly trend toward a fully mixed
to use a simple Runge-­Kutta 2, FD scheme (Materials and Methods, equilibrium state where ⟨Φ1⟩ ∕ ⟨Φ2⟩ ≈ 1. Figure 2B quantitatively
“FD discretization” section) to solve Eq. 3 and time evolve the shows that the root mean square error (RMSE) in terms of both the
MPS-­PDF. Reynolds-­averaged mean mass fraction
An initial MPS simulation (Materials and Methods, “MPS algo-
rithm” section) is performed in Fig. 1 of a purely mixing flow with- ⟨Φα ⟩ = ⟨ Φα ⟩ dx (4)
∫[0,l0 ]×3
out chemical reactions (Da = 0). The PDF is illustrated at two points

Gourianov et al., Sci. Adv. 11, eads5990 (2025) 29 January 2025 2 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Downloaded from https://www.science.org on February 07, 2025

Fig. 1. High-­dimensional PDF of a flow undergoing turbulent mixing revealed by TN simulation. Here, the Fokker-­Planck Eq. 3 is solved for a PDF f φ1 , φ2 ; x, t
( )

over chemical mass fractions φ1 , φ2, at CΩ = 1, Da = 0 in the presence of a Pe = 103 velocity field U characterized by vortices and a jet along x1 (Materials and ( Methods,)
1
“Flow case definition” section). The 5D is represented by a MPS ansatz at , on a ×5
grid and is visualized here for , 1, 12 and
( )
( ) f φ 1 , φ 2 ; x, t χ = 128 128 x ∕ l 0 = 2
x ∕ l0 = 0, 12 , 1 at times t ∕ T0 = 0,0.125,1,2 in the left and right columns, while corresponding mean mass fraction ratios ⟨Φ1⟩ ∕ ⟨Φ2⟩ are shown in the center.

and ⟨Φα⟩ decrease roughly polynomially in χ for all CΩ. MPS algorithm (Materials and Methods, “MPS algorithm” section)
Figure 2C depicts how varying the Damköhler number affects the implements a finite difference method within the MPS framework, it
accuracy of the MPS algorithm. When Da > 0, any moments must perform the MPS equivalent of operations like element-­wise,
⟨Φnα⟩ , n ∈ ℤ≥0 higher than the norm ⟨Φ0α⟩ = ⟨1⟩ can no longer be inde- matrix-­matrix and matrix-­vector multiplications, matrix and vector
pendently computed. We therefore rather look at two quantities that our additions and subtractions, and inner and outer products, in addition
simulation must preserve: the norm, which must equal unity across to MPS-­specific operations like singular values and QR decompositions
α, x, t, and the difference in consumption between the two species to enforce the maximal bond dimension and ensure the MPS stays in
⟨Φ1 ⟩ − ⟨Φ2 ⟩, which should be zero for all t due to the symmetry of Sα the numerically manageable “canonical form” (14, 15). It is (straightfor-)
and the initial conditions. The figure indicates these two quantities ward to show (39) that these MPS operations all cost O χq dlogM
becoming increasingly preserved when, again, CΩ decreases and χ in- asymptotically, with q ∈ ℤ>0 depending on the operation.
creases. Notably, the errors decrease roughly polynomially in χ. However, The element-­wise multiplication operation is the most expensive
varying Da has little impact on the accuracy. This is because the chemical at q = 4 [(39), section( 4.6], making
) the asymptotic complexity of our
reaction largely just drives the PDF in compositional space toward the scheme as a whole O χ4 dlogM per timestep. Thus, for ( Mt timesteps,
origin (as seen in Fig. 3), without significantly affecting its structure. the total cost of the time evolution will approach O Mt χ4 dlogM at
)
very large χ; although in practice for small and intermediate χ, the
Computational complexity empirical cost scales much milder (Supplementary Text, “Empirical
The maximal bond dimension χ not only sets the accuracy of the MPS computational cost” section). In comparison, standard FD
) schemes
simulation but also determines the computational cost. Because our are exponentially more expensive in d, costing O Mt M d .
(

Gourianov et al., Sci. Adv. 11, eads5990 (2025) 29 January 2025 3 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Downloaded from https://www.science.org on February 07, 2025

Fig. 2. Accuracy convergence of TN algorithm. The influence of χ, CΩ and Da on the accuracy of the MPS simulation is outlined here. (A) and (B) contrast numerically
exact means against those extracted from the MPS algorithm. In (A), the ratio ⟨Φ1⟩∕⟨Φ2⟩ is visualized at x3 ∕ l0 = 21 for times t ∕ T0 = 0.125, 0.25, 1, and 2, top to bottom.
The leftmost column corresponds to the exact solution (which can only be practically computed for Da = 0 through Eq. 9), while the next six columns come from
MPS simulations at varying χ, CΩ. The differences between the exact and MPS solutions are quantified in the lower (B) plot; the upper (B) plot shows how well the total
species amounts ⟨Φα ⟩ (see Eq. 4) are preserved through the simulation. In (C), the RMSE in two basic statistics is computed: the difference in species consumption
⟨Φ1 ⟩ − ⟨Φ2 ⟩, which should always equal zero, and the space-­averaged norm ⟨1⟩, which should always equal one. All RMSEs are mathematically defined in Materials
and Methods, “Error measures” section.

There is also the question of preparing initial states and extracting

statistics. Regarding the former, the 3D ⟨ U ⟩ (x), γSGS (x) and d-­
dimensional f φ1 , φ2 ; x, t = 0 can be computed using either the pro-
( )
longation method {see [(39), section( 4.4] and )(40)} or the tensor-­cross
algorithm (30, 41–44), both at O χ3 dlogM cost. As for the latter,
computing expectation values boils down to doing the MPS equiva-
lent of matrix-­vector multiplication and inner products, which are,
as noted previously, inexpensive and straightforward operations.
For instance, at any given ) the 3D mean ⟨Φα⟩ can be ex-
timestep,
tracted from f at O M 3 χ2 dlogM complexity, while the cost is
(

O χ2 dlogM for the scalar ⟨Φα ⟩.

( )

Integrated quantities
The satisfactory accuracy and subexponential cost of our MPS
scheme allows us to directly compute the PDF, visualize it, and ex-
tract from it all relevant integrated quantities.
Figure 3 shows the influence of mixing and chemical reactions
Fig. 3. Final PDF for various flow parameters. The PDFs at the
[ end of
( the simulation
)] on the PDF. As expected, in the absence of chemical reaction, both
1 1 1
t∕T0 = 2 are shown here in the center of the spatial domain x∕l0 = 2 , 2 , 2 for all species tend toward the fully mixed values ⟨Φα⟩ (t → ∞) → 0.5 at a
( )

combinations of CΩ , Da. The PDFs are computed using χ = 128 MPS simulations. rate governed by CΩ. Whereas in the reacting flow simulations,

Gourianov et al., Sci. Adv. 11, eads5990 (2025) 29 January 2025 4 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

χ excessively high is expensive due to the O χ4 dlogM asymptotic

( )
⟨Φα⟩ (t → ∞) → 0 at a rate that increases with CΩ and Da. Visually,
we see that increasing CΩ leads to a PDF that is more concentrated cost of the algorithm. However, for slower mixing rates, high accuracy
along φ1 = φ2 (implying a more mixed fluid), while increasing Da is achievable at very low χ even when the chemical reaction rates are
takes the PDF closer to the origin (meaning more of the reactants high. For instance, at CΩ = 0.25, Da = 1.5, the algorithm
( ) is accurate
with just χ = 32. This is equivalent to respective  106 and  103
( )
have been consumed). Multimodality is also evident in some PDFs;
this is a result of convective and diffusive transport in x-­space. factor reductions in memory and computational costs [Materials and
The trends of Fig. 3 are reflected in the integrated quantities plot- Methods, Eq. 12, and Supplementary Text, “Empirical computational
ted in Fig. 4. The first row illustrates ⟨Φ1 ⟩ going from being con- cost” section] compared to conventional FD schemes, allowing the
served at Da = 0 to being consumed at rates increasing with Da, as time evolution to be executed on a single CPU core in only a couple of
expected. The consumption also slightly increases with the SGS hours, instead of days on a supercomputer.
mixing rate. The following row shows the negative of the (Reynolds The results shown here are only an early indication of what is pos-
averaged) scalar covariance sible: there exists great scope for improvement in both the algorithm
and its implementation. For example, using tensor-­cross or other
R12 = ⟨Φ1⟩⟨Φ2⟩ − ⟨Φ1⟩ ⟨Φ2⟩ (5) algorithms (45) to perform element-­wise multiplications might reduce
which decays in t due to ⟨Φ1⟩∕⟨Φ2⟩ approaching unity as the flow be- the complexity of our scheme to ∼χ3 without significantly sacrificing
comes increasingly mixed. Last, the last row exhibits the covariance accuracy. Furthermore, better optimized software running on special-
ized computing architectures will allow for much larger bond dimen-
Y12 = ⟨Φ1⟩⟨Φ2⟩ − ⟨Φ1⟩ ⟨Φ2⟩ (6) sions and system sizes: We are currently simulating a M d = 1285 = 235
grid at χ = 128, while the current record is a quantum physics simula-
At initial times, −Y12 increases due to large gradients in ⟨Φα⟩,
tion on a grid equivalent of M d = 2400 at χ = 32768, performed on a
followed by a decrease due to mixing((with higher) mixing-­rates lead-

Downloaded from https://www.science.org on February 07, 2025

tensor processing unit pod (46).
ing to a faster decay). The ratio R12 ∕ R12 + Y12 is consistently above We decided to use an MPS ansatz because it closely matches the
one half, implying that most of the energy of the eddies is resolved structure of the PDF for this particular flow; other flows may have dif-
during the simulations. The statistical trends observed in Fig. 4 are ferent PDFs for which alternative ansatze could be better suited (47).
consistent with those reported in turbulence literature (35). Fortunately, there exists a rich and growing selection of TNs to pick
from, each carrying their advantages and disadvantages. These range
from 2D generalizations of MPSs (48), hierarchical networks (49, 50),
DISCUSSION and even networks that might someday leverage quantum hardware
The results imply that MPSs are able(to efficiently ) exploit structure (51). There is also the exciting prospect of catering the TN ansatz to
within turbulence PDFs. The PDF f φ1 , φ2 ; x, t of our 3D chemi- the structure of the PDF in an automated manner (52). Typically, more
cally reactive flow case [Eq. 3] is of an orderly shape, and the cou- complex ansatze are able to encode solutions with higher accuracy at
pling between its dimensions is limited by the SGS mixing rate CΩ. lower χ but are costlier to manipulate. Balancing such considerations
Exploiting these structures permits our MPS scheme to accurately while exploring alternative TN geometries for probabilistic turbulence
and efficiently represent f and evolve it through time. In the future, simulations is a promising avenue of future investigation.
a more realistic model should be considered where ( the velocity ) field Turbulence is just one example of a complex system; there are
is also included within the PDF, turning f = f u, φ1 , φ2 ; x, t into a many others, ranging from biological organisms to financial markets
d = 8 + 1 dimensional object. (53). These kind of systems exhibit chaotic and unpredictable dynam-
Ensuring the MPS algorithm maintains both accuracy and effi- ics that ultimately require statistical descriptions (54). The most fun-
ciency requires carefully selecting χ (see figs. S1 to S3). Figure 2 indi- damental way of doing so is by modeling their PDFs. Yet, such PDFs
cates that varying the Damköhler number does not significantly affect are typically prohibitively high dimensional (as displayed here for the
( ) the SGS mixing rate requires χ to in
the accuracy, while increasing case of turbulence), which has made solving their governing Fokker-­
turn increase as χ ∼ poly CΩ for accuracy to be maintained. Setting Planck equations infeasible, until now. This work is a first demonstra-
tion in how the problem can be overcome via a simple TN. More
advanced TN ansatze and algorithms will be developed in time, hold-
ing the promise of enabling large-­scale probabilistic simulations both
within the field of fluid dynamics and beyond.

MATERIALS AND METHODS

Flow case definition
In Eq. 3, the mean velocity field ⟨ U ⟩ is set a priori. To ensure adequate
convective mixing and for the flow to be interesting, we elected to set
the velocity field to a jet moving through a Taylor-­Green vortex

Fig. 4. Statistics extracted from PDF for different flow parameters. In the first (x2 ∕l0 −1∕2)2 + (x3 ∕l0 −1∕2)2
−
⟨U1 ⟩∕u0 = coskx1 sinkx2 sinkx3 − e 2(1∕6)2
row, the total amount of the first species is plotted. The next row is of the Reynolds-­
averaged covariance R12, while the final row quantifies the space-­averaged covari- ⟨U2 ⟩∕u0 = sinkx1 coskx2 sinkx3 , (7)
ance Y12 . These quantities are defined in Eqs. 4 to 6. The PDFs are computed at
χ = 128. ⟨U3 ⟩∕u0 = − 2sinkx1 sinkx2 coskx3

Gourianov et al., Sci. Adv. 11, eads5990 (2025) 29 January 2025 5 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Here, the vortex wavenumber k is set to k = 4π ∕ l0. Particular care must be taken when defining the boundary con-
The initial (t = 0) PDF is chosen to be a Gaussian step function ditions for this problem. While in x-­space, one may simply assume
periodic boundaries, in compositional space, the boundary condi-
⎧ − (φ1 −3∕4)2 + (φ2 2 −1∕4)2 1 3 tions must be defined in a way that stops probability leaking out of
1 ⎪e 2(1∕8) , ≤ x1 < ,
the domain. This is achieved by making the composition space
f (t = 0) = 4 4 (8)
2π(1∕8)2 ⎪ − (φ1 −1∕4)2 + (φ2 −3∕4)2 ghosts points for any order-­n discrete derivative of f follow
⎨
2
⎩e 2(1∕8) , otherwise
M−1 [
∑ Δn f
]
that has undergone numerical smoothing in the x dimensions (the =0 (11)
i=0
Δφnα i
smoothing is meant to soften the step function at the x1 = 41 , 43
boundary sufficiently to avoid numerical instabilities during time with i denoting a discretized (equidistantly distributed) lattice point.
evolution). The initial PDF is illustrated in the first rows of Fig. 1. Equation 11 imposes f−1 = −f0 & fM = −fM−1 for the first derivative, and
The M = 128, d = 5 grid is sufficient for the simulation to be con- f−1 = f0 & fM = fM−1 for the second, under our CFD2 discretization.
ducted with parameters that make physical sense: we set Cs = 0.11,
Δ 𝓁 = 3Δx = 3l0 ∕ M , Pe = 103, CΩ ∈ [0.25,1], and Da ∈ [0,1.5]. MPS algorithm
Our MPS algorithm implements the aforementioned RK2-­ CFD2
Moment equations scheme on the MPS manifold (14, 55). This entails parameterizing all
The zeroth moment of the Fokker-­Planck Eq. 3 recovers the hydro- the vectors (like f , ⟨ Ui ⟩, and ⟨Φα⟩) in Eq. 10 as MPSs [(39), section
dynamic continuity equation ∇ ⋅ ⟨ U ⟩ = 0, while the first gives an 3.4], and the matrices (e.g., Δ ∕ Δxi) as analogous matrix product
equation for the mean mass fractions ⟨Φα⟩ (x, t) operators [(39), section 3.5]. Then, within the MPS format, the time-­

Downloaded from https://www.science.org on February 07, 2025

� � stepping is performed in a standard manner using the arithmetic op-
𝜕 ⟨Φα ⟩ 𝜕 � � 𝜕 ⟨Φα ⟩
+ ⟨Ui ⟩ α = (9) erations outlined in the “Computational complexity” section.
⟨Φ ⟩
γ + γSGS + ⟨Sα ⟩
𝜕t 𝜕 xi 𝜕 xi 𝜕 xi It is essential to control the bond dimension during the MPS simu-
In nonreactive flows (Da = 0), S = 0 and Eq. 9 can be cheaply and lation. The arithmetic operations that time-­evolve f lead to its bond
accurately solved using a standard FD scheme to obtain a “numeri- dimension growing exponentially in time, if not truncated [(39), sec-
cally exact” ⟨Φα⟩ solution (in the sense that there is no truncation tion 3.5.3]. In our code, we use the singular values decomposition to
error in χ, as explained in Materials and Methods, “Error measures” truncate the bond dimension of f such that it is always limited to χ. As
section). This is used to check the accuracy of the MPS algorithm in for the other vectors and matrices, these objects remain constant in
Fig. 2 (A and B). It is not possible to obtain a numerically exact ⟨Φα⟩ time and their bond dimensions are all of order (10).
when Da > 0, because a closure model would be required for ⟨Sα⟩. The maximal bond dimension χ defines the NVPP. For an MPS
representation of f , the number of parameters becomes
FD discretization
The simulations are performed on equidistant Cartesian grids with N
∑ N−1
∑
M = 128 gridpoints along each dimension. The derivatives in Eqs. 3 NVPP = 2 p(n − 1)p(n) − p(n)2 (12)
and 9 are discretized in a simple manner: the temporal derivative n=1 n=1

with an explicit Runge-­Kutta 2 scheme and a second-­order-­accurate with N = log2 M d, (M must be a power of 2) )being the number of ten-
central FDs (CFD2) discretization of the x, φ1 , φ2 derivatives. sors in the MPS, and p(n) = min 2n , 2N−n , χ being the size of the nth
(
However, discretizing Eq. (3) creates the practical challenge of bond of the MPS. The first sum gives the total number of parameters in
handling delta-­functions. The LMSE model forces each Φα toward the MPS, while the second sum represents the intrinsic gauge degrees of
⟨Φα⟩ at every x, t , equivalent to the PDF in composition space mov- freedom of the MPS format (55). When χ is maximal, i.e., χ = 2⌊N∕2⌋, we
ing toward a delta function centered around the mean of the mass get NVPP = 2N = M d and that f is represented exactly on the M ×d grid.
fractions. Resolving delta functions on discretized grids is difficult,
as their sharp gradients reduce the accuracy and stability of any nu- Error measures
merical scheme used to compute the PDF transport. While often The errors in Fig. 2 (B and C) are computed using the RMSE mea-
this is dealt with by using highly dissipative discretizations of de- sure across χ. In the first figure, the upper Error2b↑ and lower Error2b↓
rivatives (e.g., upwinding), we rather choose to simply modify the are computed by averaging the spatially averaged mean quantities
LMSE model in Eq. 3 through the addition of an artificial dissipa- across α, t and α, t, x, respectively. In Fig. 2C, the averaging is done
tion term to the compositional space. Doing this while discretizing across just t and t, x to compute Error2c↑, Error2c↓. Mathematically,
the Fokker-­Planck PDE results in these errors can be expressed as
�
Δf Δf
�
Δ � � Δf
� 1 � � �
+ ⟨Ui ⟩ − γ + γSGS = Error2b↑ (χ) = ℰ t ⟨Φα ⟩(χ), 1∕2 ,
Δt Δxi Δxi Δx 2 α=1,2
� � i (10)
Δf
�
Δ � � Δ � � 1 �
Ω φ − ⟨Φα ⟩ f + CΩ μ − S f
� �
Δφα mix α Δφα Δφα α Error2b↓ (χ) = ℰ t,x ⟨Φα ⟩(χ), ⟨Φα ⟩(exact)
2 α=1,2 (13)
with the artificial dissipation governed by μ. This parameter needs to � � �
be set to be as small as possible to minimally affect the accuracy, Error2c↑ (χ) = ℰ t ⟨Φ1 ⟩(χ) − ⟨Φ2 ⟩(χ), 0 ,
while still being large enough to ensure f is well resolved on M. From � � �
trial and error, we find μ = 4 ⋅ 10−3 l 0 works well for M = 128.
u Error2c↓ (χ) = ℰ t,x ⟨1⟩(χ), 1
0

Gourianov et al., Sci. Adv. 11, eads5990 (2025) 29 January 2025 6 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

with ⟨Φα⟩ (χ) being extracted from the MPS-­PDF solution of Eq. 10 19. S. Trotzky, Y.-­A. Chen, A. Flesch, I. P. McCulloch, U. Schollwöck, J. Eisert, I. Bloch, Probing
the relaxation towards equilibrium in an isolated strongly correlated one-­dimensional
while ⟨Φα⟩ (exact) is the solution found by directly solving Eq. 9 with a
Bose gas. Nat. Phys. 8, 325–330 (2012).
standard RK2-­CFD2 FD scheme; this solution is “numerically exact” in 20. B.-­X. Zheng, C.-­M. Chung, P. Corboz, G. Ehlers, M.-­P. Qin, R. M. Noack, H. Shi, S. R. White,
the sense that it does not suffer from any truncation error in χ [although S. Zhang, G. K.-­L. Chan, Stripe order in the underdoped region of the two-­dimensional
a truncation error from the FD discretization itself remains, this error is Hubbard model. Science 358, 1155–1160 (2017).
slight due to the smoothness of ⟨Φα⟩ (exact)]. The functions 21. C. Huang, F. Zhang, M. Newman, X. Ni, D. Ding, J. Cai, X. Gao, T. Wang, F. Wu, G. Zhang,
H.-­S. Ku, Z. Tian, J. Wu, H. Xu, H. Yu, B. Yuan, M. Szegedy, Y. Shi, H.-­H. Zhao, C. Deng, J. Chen,
Efficient parallelization of tensor network contraction for simulating quantum
( ) dx [ ]
computation. Nat. Comput. Sci. 1, 578–587 (2021).
Et,x g, g0 = E g(x), g0 (x) ,
∫[0,l0 ]×3 l 3 t 22. Y. Zhou, E. M. Stoudenmire, X. Waintal, What limits the simulation of quantum
0
computers? Phys. Rev. X 10, 041038 (2020).
2T0 (14) 23. J. Tindall, M. Fishman, E. M. Stoudenmire, D. Sels, Efficient tensor network simulation of
( ) dt [ ]2
IBM’s Eagle kicked ising experiment. PRX Quantum 5, 010308 (2024).
Et g, g0 = g(t) − g0 (t)
∫ 2T0 24. T. Begušić, J. Gray, G. K.-­L. Chan, Fast and converged classical simulations of evidence for
0 the utility of quantum computing before fault tolerance. Sci. Adv. 10, eadk4321 (2024).
implement temporal (Et ) and space-­time (Et,x ) averaging. 25. N. Gourianov, M. Lubasch, S. Dolgov, Q. Y. van den Berg, H. Babaee, P. Givi, M. Kiffner,
Note that since both time and space are discretized during the simu- D. Jaksch, A quantum-­inspired approach to exploit turbulence structures. Nat. Comput.
Sci. 2, 30–37 (2022).
lations, the above integrals are both performed numerically using a 26. E. Ye, N. F. Loureiro, Quantum-­inspired method for solving the Vlasov-­Poisson equations.
simple step quadrature. In space, the integrals are computed using all Phys. Rev. E 106, 035208 (2022).
the M = 128 gridpoints along each dimension. In time, the integral is 27. E. Ye, N. F. Loureiro, Quantized tensor networks for solving the Vlasov-­Maxwell equations.
T T 3T arXiv:2311.07756 [physics.comp-­ph] (2024).
computed over the 17 time samples t = 0, 80 , 40 , 8 0 ,…, 2T0.
28. R. D. Peddinti, S. Pisoni, A. Marini, P. Lott, H. Argentieri, E. Tiunov, L. Aolita, Quantum-­

Downloaded from https://www.science.org on February 07, 2025

inspired framework for computational fluid dynamics. Commun. Phys. 7, 135 (2024).
29. L. Hölscher, P. Rao, L. Müller, J. Klepsch, A. Luckow, T. Stollenwerk, F. K. Wilhelm,
Supplementary Materials Quantum-­inspired fluid simulation of 2D turbulence with GPU acceleration.
This PDF file includes: arXiv:2406.17823 [physics.flu-­dyn] (2024).
Supplementary Text 30. M. Ritter, Y. Núñez Fernández, M. Wallerberger, J. von Delft, H. Shinaoka, X. Waintal,
Figs. S1 to S3 Quantics tensor cross interpolation for high-­resolution parsimonious representations of
References multivariate functions. Phys. Rev. Lett. 132, 056501 (2024).
31. S. Rohshap, M. K. Ritter, H. Shinaoka, J. von Delft, M. Wallerberger, A. Kauch, Two-­particle
calculations with quantics tensor trains – Solving the parquet equations.
REFERENCES AND NOTES arXiv:2410.22975 [cond-­mat.str-­el] (2024).
1. E. Hopf, Statistical hydromechanics and functional calculus. J. Rat. Mech. Anal. 1, 32. F. A. Jaberi, R. S. Miller, C. K. Madnia, P. Givi, Non-­Gaussian scalar statistics in
87–123 (1952). homogeneous turbulence. J. Fluid Mech. 313, 241–282 (1996).
2. A. S. Monin, A. M. Yaglom, Statistical Fluid Mechanics (MIT Press, Cambridge, MA) (1975). 33. J. H. Chen, Petascale direct numerical simulation of turbulent combustion–Fundamental
3. S. B. Pope, Turbulent Flows (Cambridge Univ. Press, Cambridge, UK) (2012). insights toward predictive models. Proc. Combust. Inst. 33, 99–123 (2011).
4. R. O. Fox, Computational Models for Turbulent Reacting Flows (Cambridge Univ. Press, 34. A. G. Nouri, M. B. Nik, P. Givi, D. Livescu, S. B. Pope, Self-­contained filtered density
Cambridge, UK) (2009). function. Phys. Rev. Fluids 2, 094603 (2017).
5. C. Dopazo, Recent Developments in PDF Methods, in Turbulent Reacting Flows, P. A. Libby, 35. P. Givi, Filtered density function for subgrid scale modeling of turbulent combustion.
F. A. Williams, Eds. (Academic Press, London, England), chap. 7, pp. 375–474 (1994). AIAA J. 44, 16–23 (2006).
6. F. A. Williams, Combustion Theory (The Benjamin/Cummings Publishing Company, Menlo 36. J. Smagorinsky, General circulation experiments with the primitive equations. I. The basic
Park, CA), Ed. 2. (1985). experiment. Monthly Weather Rev. 91, 99–164 (1963).
7. D. Livescu, A. G. Nouri, F. Battaglia, P. Givi, Eds. Modeling and Simulation of Turbulent 37. E. E. O’Brien, The probability density function (PDF) approach to reacting turbulent flows,
Mixing and Reaction: For Power, Energy and Flight (Springer, Germany) (2020). in Turbulent Reacting Flows, P. A. Libby, F. A. Williams, Eds. (Springer-­Verlag, Heidelberg),
8. S. B. Pope, PDF methods for turbulent reactive flows. Prog. Energy Combust. Sci. 11, vol. 44 of Topics in Applied Physics, chap. 5, pp. 185–218 (1980).
119–192 (1985). 38. M. Kiffner, D. Jaksch, Tensor network reduced order models for wall-­bounded flows. Phys.
9. V. Hiremath, S. R. Lantz, H. Wang, S. B. Pope, Computationally-­efficient and scalable Rev. Fluids 8, 124101 (2023).
parallel implementation of chemistry in simulations of turbulent combustion. Combust. 39. N. Gourianov, Exploiting the structure of turbulence with tensor networks, Ph.D. thesis,
Flame 159, 3096–3109 (2012). University of Oxford (2022).
10. H. Zhou, P. Givi, Z. Ren, Filtered density function: A stochastic closure for coarse grained 40. M. Lubasch, P. Moinier, D. Jaksch, Multigrid renormalization. J. Comput. Phys. 372,
simulation, in Coarse Graining Turbulence: Modeling and Data-­Driven Approaches and their 587–602 (2018).
Applications (Cambridge Univ. Press), chap. 4 (2025). 41. I. Oseledets, E. Tyrtyshnikov, TT-­cross approximation for multidimensional arrays. Linear
11. D. Poulin, A. Qarry, R. Somma, F. Verstraete, Quantum simulation of time-­dependent Algebra Appl. 432, 70–88 (2010).
Hamiltonians and the convenient illusion of Hilbert space. Phys. Rev. Lett. 106, 42. S. Dolgov, D. Savostyanov, Parallel cross interpolation for high-­precision calculation of
170501 (2011). high-­dimensional integrals. Comput. Phys. Commun. 246, 106869 (2020).
12. S. R. White, Density matrix formulation for quantum renormalization groups. Phys. Rev. 43. B. Ghahremani, H. Babaee, A DEIM Tucker tensor cross algorithm and its application to
Lett. 69, 2863–2866 (1992). dynamical low-­rank approximation. Comput. Methods Appl. Mech. Eng. 423, 116879
13. G. Vidal, Efficient classical simulation of slightly entangled quantum computations. Phys. (2024).
Rev. Lett. 91, 147902 (2003). 44. Y. Núñez Fernández, M. K. Ritter, M. Jeannin, J.-­W. Li, T. Kloss, T. Louvet, S. Terasaki,
14. U. Schollwöck, The density-­matrix renormalization group in the age of matrix product O. Parcollet, J. von Delft, H. Shinaoka, X. Waintal, Learning tensor networks with tensor cross
states. Ann. Phys. 326, 96–192 (2011). interpolation: New algorithms and libraries. arXiv:2407.02454 [physics.comp-­ph] (2024).
15. R. Orús, Tensor networks for complex quantum systems. Nat. Rev. Phys. 1, 538–550 (2019). 45. A. A. Michailidis, eC. Fenton, M. Kiffner, Tensor train multiplication. arXiv:2410.19747
16. S. R. Clark, D. Jaksch, Dynamics of the superfluid to Mott-­insulator transition in one [physics.comp-­ph] (2024).
dimension. Phys. Rev. A 70, 043612 (2004). 46. M. Ganahl, J. Beall, M. Hauru, A. G. M. Lewis, T. Wojno, J. H. Yoo, Y. Zou, G. Vidal, Density
17. A. Feiguin, S. Trebst, A. W. W. Ludwig, M. Troyer, A. Kitaev, Z. Wang, M. H. Freedman, matrix renormalization group with tensor processing units. PRX Quantum 4, 010317 (2023).
Interacting anyons in topological quantum liquids: The golden chain. Phys. Rev. Lett. 98, 47. I. Glasser, R. Sweke, N. Pancotti, J. Eisert, J. I. Cirac, Expressive power of tensor-­network
160409 (2007). factorizations for probabilistic modeling. Adv. Neural Inf. Process. Syst. 32, 1496–1508 (2019).
18. M. Cheneau, P. Barmettler, D. Poletti, M. Endres, P. Schauß, T. Fukuhara, C. Gross, I. Bloch, 48. F. Verstraete, M. M. Wolf, D. Perez-­Garcia, J. I. Cirac, Criticality, the area law, and the
C. Kollath, S. Kuhr, Light-­cone-­like spreading of correlations in a quantum many-­body computational power of projected entangled pair states. Phys. Rev. Lett. 96,
system. Nature 481, 484–487 (2012). 220601 (2006).

Gourianov et al., Sci. Adv. 11, eads5990 (2025) 29 January 2025 7 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

49. V. Murg, F. Verstraete, O. Legeza, R. M. Noack, Simulating strongly correlated quantum acknowledges support by US-­AFSOR grant FA8655-­22-­1-­7027 and the UKRI “Quantum
systems with tree tensor networks. Phys. Rev. B 82, 205105 (2010). Computing and Simulation Hub” grant EP/P009565/1. P.G. acknowledges support by
50. G. Evenbly, G. Vidal, Class of highly entangled many-­body states that can be efficiently US-­AFOSR grant FA9550-­23-­1-­0014. D.J. acknowledges support by the European Union’s
simulated. Phys. Rev. Lett. 112, 240502 (2014). Horizon Programme (HORIZON-­CL42021-­DIGITALEMERGING-­02-­10) grant agreement
51. D. Jaksch, P. Givi, A. J. Daley, T. Rung, Variational quantum algorithms for computational 101080085 QCFD, the Cluster of Excellence “Advanced Imaging of Matter” of the Deutsche
fluid dynamics. AIAA Journal 61, 1885–1894 (2023). Forschungsgemeinschaft (DFG), EXC 2056, project ID 390715994, and the Hamburg Quantum
52. R. Peng, J. Gray, G. K.-­L. Chan, Arithmetic circuit tensor networks, multivariable function Computing Initiative (HQIC) project EFRE. The project is cofinanced by ERDF of the European
representation, and high-­dimensional integration. Phys. Rev. Res. 5, 013156 (2023). Union and by “Fonds of the Hamburg Ministry of Science, Research, Equalities and Districts
53. A. Favre, H. Guitton, J. Guitton, A. Lichnerowicz, Chaos and Determinism: Turbulence as a (BWFGB)”. Author contributions: N.G. and P.G. conceived and planned the research project.
Paradigm for Complex Systems Converging Toward Final States (Johns Hopkins Univ. Press) N.G. formulated the MPS algorithm, did the analytical calculations, and wrote the software. The
(1995). numerical experiments were jointly designed by all the authors and executed by N.G. The
54. D. Bandak, A. A. Mailybaev, G. L. Eyink, N. Goldenfeld, Spontaneous stochasticity amplifies numerical results were analyzed and interpreted jointly by all the authors. N.G. and P.G. wrote
even thermal noise to the largest scales of turbulence in a few eddy turnover times. Phys. the manuscript, with contributions from the other authors. The project was supervised by P.G.
Rev. Lett. 132, 104002 (2024). and S.B.P. Competing interests: The authors declare that they have no competing interests.
55. S. Holtz, T. Rohwedder, R. Schneider, On manifolds of tensors of fixed TT-­rank. Numer. Data and materials availability: All software and data accompanying this manuscript are
Math. 120, 701–731 (2012). publicly available at https://github.com/nikitn2/tendeq and https://doi.org/10.5281/
56. J. Gray, quimb: A Python library for quantum information and many-­body calculations. zenodo.14223424. All data needed to evaluate the conclusions of the paper are present in the
J. Open Source Softw. 3, 819 (2018). paper and/or the Supplementary Materials.

Acknowledgments: We thank A. Louis at the University of Oxford for discussions and would Submitted 20 August 2024
like to acknowledge the use of the University of Oxford Advanced Research Computing (ARC) Accepted 27 December 2024
facility in carrying out this work (http://dx.doi.org/10.5281/zenodo.22558). For the purpose of Published 29 January 2025
open access, a CC BY public copyright license is applied to this manuscript. Funding: N.G. 10.1126/sciadv.ads5990

Downloaded from https://www.science.org on February 07, 2025

Gourianov et al., Sci. Adv. 11, eads5990 (2025) 29 January 2025 8 of 8