Inverse Evolution Data Augmentation for Neural PDE Solvers

Carola-Bibiane Sch\"onlieb; Chaoyu Liu; Chris Budd

REVIEW 4 major objections 4 minor 30 references

Inverse Evolution Data Augmentation for Neural PDE Solvers

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper establishes that one explicit backward-time step, after time reversal, yields data pairs satisfying implicit forward schemes, and shows that adding these pairs improves neural operators on Burgers, Allen-Cahn, and Navier-Stokes…

desk verdict Clever augmentation idea with a correct first-order identity, but the headline gain is confounded because augmented runs double the training set. read the letter →

arxiv 2501.14604 v1 pith:F2ZDS4OK submitted 2025-01-24 cs.LG

Chaoyu Liu , Chris Budd , Carola-Bibiane Schönlieb This is my paper · ORCID

classification cs.LG MSC 68T0735Q3565N99

keywords neuraloperatorsdataaugmentationinverseevolutionpartialdifferentialequationshigh-orderschemesFourierOperatorUNet

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a backward-time trick can produce abundant training data for neural operators that solve evolution equations, at a fraction of the cost of running forward solvers. The trick is to take a later-time state, advance it one explicit step of the inverse equation $u_t=-F(u)$, and then reverse the pair; the reversed pair satisfies the implicit Backward Euler scheme of the forward equation. The paper proves the same for second- and third-order Taylor-based inverse schemes, adds a mixing-and-shifting initialization step to create varied starting states, and tests the generated pairs by training Fourier Neural Operators and UNets on Burgers, Allen-Cahn, and Navier-Stokes equations. The reported result is lower test error and better robustness to noisy inputs, with the largest gains when the original dataset is small.

What carries the argument

The carrying mechanism is the reversal identity: take one explicit backward step $U^{n+1}=U^n-\Delta t\,F(U^n)$ and then swap the pair so $(V^n,V^{n+1})=(U^{n+1},U^n)$; substitution gives $(V^{n+1}-V^n)/\Delta t = F(V^{n+1})$, which is exactly the implicit Backward Euler scheme of the forward equation. This identity turns an unstable explicit backward computation into a legitimate implicit-scheme training pair. The paper's high-order variants repeat the same argument with Taylor-expansion inverse steps, yielding data that satisfy higher-order implicit schemes, while the randomized mixing initialization Eq. (7) and the rescaling preprocessing in Eq. (14) keep the generated inputs within a useful neighborhood of the solution space.

What would settle it

Take a fixed set of solution snapshots, build initializations with progressively larger mixing constants in Eq. (7), run one third-order inverse step, and compare the time-reversed pair against a high-accuracy forward reference over the same interval; if the relative $L^2$ error grows with the mixing constants or exceeds the error of the forward solver used to create the original data, the generated pairs are not reliable implicit-scheme solutions.

Watch

Extended reading notes

Core claim

The central claim is that inverse evolution is a reliable and cheap data source for learning evolution operators. If $(U^n, U^{n+1})$ is computed by one explicit Forward Euler step of the inverse equation, then the time-reversed pair $(V^n,V^{n+1})=(U^{n+1},U^n)$ satisfies the implicit Backward Euler formula $(V^{n+1}-V^n)/\Delta t = F(V^{n+1})$ of the original forward equation. The paper extends this identity to Taylor-expansion inverse schemes up to third order, claiming the generated pairs satisfy the corresponding high-order implicit schemes, so that explicit computation with relatively large time steps yields data normally associated with stable implicit solvers. The initialization combines random permutations of existing solution snapshots with constants, Eq. (7), and a normalization/rescaling preprocessing step handles sharp interfaces. Experiments on three evolution equations show that adding these pairs to the training set improves FNO accuracy in all tested settings and improves UNet in most, with particularly clear gains at 1000-pair datasets.

Load-bearing premise

The method's reliability rests on the assumption that the randomly mixed and shifted snapshots in Eq. (7) are close enough to genuine solutions of the evolution equation that one explicit backward step, after time reversal, still lands near a true forward trajectory.

Editorial extensions

If this is right

Every pair produced by one explicit inverse step of the form (4) is, after reversing the time order, a Backward Euler pair of the forward equation, so the augmented data inherit the stability and consistency of an implicit scheme without solving nonlinear equations.
The high-order Taylor inverse schemes (8)-(10) yield pairs that satisfy the corresponding high-order implicit schemes, which the accuracy tests show reduces the relative $L^2$ error of generated data at moderate $\Delta t$.
The method reduces the cost of training-data generation: for 100 pairs at resolution 256, the measured time falls from about 3.2 s to 0.27 s for 1D Burgers, from 11.8 s to 0.15 s for 2D Allen-Cahn, and from 74.8 s to 0.033 s for Navier-Stokes.
FNO trained on augmented data achieves lower test error than on original data alone across Burgers, Allen-Cahn, and Navier-Stokes settings, with the biggest relative gains at 1000 training pairs; the Allen-Cahn noise experiments indicate the augmented training also improves robustness to input perturbations.
UNet also benefits in most settings, though less than FNO, consistent with the paper's suggestion that architectures with larger capacity extract more from the more complex inputs produced by inverse evolution.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Testable extension: because the reversal identity is algebraic, the same augmentation could be applied to any architecture that learns time-stepping maps, including Transformers or graph-based operators; the measured gain would likely grow with the model's ability to represent high-frequency features, paralleling the FNO-versus-UNet gap reported here.
The initialization in Eq. (7) is essentially a linear mixing of solution snapshots, so the central unresolved risk is distribution shift; a direct test would be to measure the distance between generated inputs and the true solution manifold as the mixing constants $\lambda_j$ and $C$ vary, and to compare against training on randomly perturbed snapshots of equal size.
The paper's robustness result for Allen-Cahn suggests the high-frequency content in augmented inputs acts as a regularizer; a natural way to check this would be to compare the spectral bias or the norm of the learned operator's Jacobian with and without augmented data.
Using pseudo-spectral spatial discretization for the inverse step, which the paper already finds less sensitive to $C$ for Navier-Stokes, could be combined with adaptive time stepping to push the method into regimes where explicit high-order inverse steps currently diverge.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the inverse-evolution swap is a genuinely neat trick. The statement that one explicit backward step followed by time reversal yields a pair satisfying implicit Euler is elementary and correct. Using that as a data generator for neural operators, with random-combination initialization and Taylor-based high-order variants, is new relative to the cited augmentation literature (Lie point symmetry, general covariance). Self-citation to [20] is appropriate; the augmentation pipeline is distinct from the regularizer there.

What the paper does well: the method is cheap (Table 2), the generated data are checked against true solutions in Table 1, and the authors are honest about limitations\u2014instability for chaotic or sharply interfaced problems and the restriction of preprocessed data to a specific space. The FNO gains at 1000 pairs are sometimes large (Allen-Cahn from 4.1e-2 to 5.2e-3), which is not nothing.

The soft spots are real, and one is load-bearing. Section 3.3 states that augmented runs use 2N training pairs when baselines use N. Tables 3 and 4 therefore compare one dataset size against twice that size. Any benefit is compatible with the trivial hypothesis that more data helps; the inverse-evolution mechanism is never isolated. That is not a minor issue\u2014it is the central empirical claim. The paper needs a matched-size control (for example, N original + N additional original, or N original + N pairs generated by a non-physics random process) before the efficacy conclusion is supported. Second, the claim that high-order Taylor pairs \u201ccan be proved to satisfy the corresponding high-order implicit schemes\u201d is asserted without proof. It may be true, but it is not shown, and Table 1 shows the high-order schemes degrade badly at larger time steps. Third, Equation (7) random combinations are asserted to stay near the solution manifold without any quantitative check; Figure 3 shows some generated inputs look plausible, but there is no distribution-shift measure. Minor: no code, seeds, or error bars, so the magnitude of the gains is hard to assess.

On balance, the paper is worth engaging. The mathematical core is sound at first order, the method is cheap, and the confound is fixable with one additional experiment. I would send it to review with a request for matched-size comparisons and code, but I would not accept the efficacy claim as it stands.

Referee Report

4 major / 4 minor

Summary. The paper proposes a data augmentation strategy for training neural operators on evolution equations. The idea is to initialize a backward-time integration from randomized combinations of existing solution snapshots, take one explicit (or Taylor-based high-order) step of the inverse equation, and then reverse the pair so that it formally satisfies an implicit scheme of the forward equation. The authors validate the accuracy of generated pairs on heat, Burgers, Allen-Cahn, and Navier-Stokes equations and report that adding such pairs improves FNO and UNet test errors (Tables 3 and 4).

Significance. If the method worked as claimed, it would be a useful and cheap way to generate training pairs for neural PDE solvers without solving implicit systems. The first-order equivalence is simple and correct, and the computational-time comparison in Table 2 is attractive. However, the experimental evidence does not isolate the proposed mechanism: every augmented run contains twice as many training pairs as the corresponding baseline run. The high-order equivalence is also stated without proof. These gaps are central to the paper's claims, so the contribution is not yet established.

major comments (4)

[Section 3.3 and 3.4, Tables 3 and 4] The reported gains are confounded with dataset size. The text states that the augmented datasets contain the original pairs plus an equal number of generated pairs, so the 'yes' rows train on 2000, 10000, and 20000 pairs while the 'no' rows train on 1000, 5000, and 10000 pairs. Any improvement in the augmented columns is therefore compatible with the trivial hypothesis that doubling the number of training examples helps, independent of the inverse-evolution construction. To support the claim that inverse evolution itself improves performance, the authors should add a matched-data-size control: either train baselines on the same total number of original pairs, or train augmented models on a random subsample of the full augmented set of the same size as the baseline. Without such a control, the central empirical claim of the paper is not established.
[Section 2, High-order Schemes (after Eq. (13))] The statement that data pairs from the Taylor-based high-order inverse schemes 'can be proved' to satisfy the corresponding high-order implicit schemes is asserted without proof, and no reference is given. This claim is load-bearing because the main FNO experiments use the third-order scheme. The authors should either provide the explicit high-order implicit relations (for the second- and third-order cases) or state the precise approximation property that the reversed pairs satisfy. As written, the theoretical justification for the high-order augmentation is missing.
[Section 2, Eq. (7)] The initialization via U*_i = sum_j lambda_j U_{R_j(i)} + C is a random linear combination of shuffled solution snapshots plus a constant, but the paper gives no evidence that such initializations lie close to the solution manifold of the evolution equation. The assertion that nonlinearity makes the generated pairs distinct from the original data does not establish that the pairs are physically relevant or that one inverse-evolution step maps them to states on the solution manifold. A concrete check is needed, such as measuring the distribution shift between generated inputs and test-time inputs, or reporting the residual of the generated pairs under a high-accuracy reference solver. Without this, the generated data may be out-of-distribution, and the experimental gains may be due to a broadened input distribution rather than to the inverse-evolution mechanism.
[Section 3.1 vs. Section 3.3] The accuracy of the generated data is not reported for the exact settings used in the neural-operator experiments. Table 1 covers Allen-Cahn with epsilon=0.05 and Navier-Stokes with nu=0.001, but the FNO experiments also use Allen-Cahn epsilon=0.01 and Navier-Stokes nu=0.0001; Burgers accuracy is reported only for nu=0.1, while the FNO experiments use nu=0.001 as well. Since the reliability of the augmented data is the premise of the method, the authors should report the generated-data accuracy for those configurations, or clearly state which settings in Table 1 correspond to the data actually added in Tables 3 and 4.

minor comments (4)

[Section 2, Eq. (5)] The definition of (V^n, V^{n+1}) := (U^{n+1}, U^n) is notationally confusing because the superscripts imply a forward time step while the pair is time-reversed; please define the time levels explicitly (for example, V^n = U^{n+1} and V^{n+1} = U^n) to avoid apparent inconsistency.
[Table 1] The 'Magnitude' row in Table 1 is ambiguous: it is unclear whether the scaling factor applies to all entries in the column section or only to the first row, and the caption does not define the em dash placeholder. Please clarify by using scientific notation in each cell and defining '—' in the caption.
[Section 3.3] Hyperparameters for Eq. (7) are incompletely specified: the choice of lambda_j and C is given only for Burgers nu=0.001 (C in [-0.1, 0.1]), while the experiments for Allen-Cahn and Navier-Stokes do not state the values or ranges used, which prevents reproduction.
[Section 2 and Figure 2] There are several typographical errors, including 'generated reliable data' in Section 2 and 'H 2 norm' in Section 2, which should read 'H^2 norm'; additionally, the caption and Eq. (14) do not specify how the rescaling coefficient a is chosen for the experiments.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the inverse-evolution data-pair identity is derived from elementary time-reversal algebra, and the only self-citation is historical provenance.

full rationale

The central derivation is self-contained. Equations (4)-(6) show that reversing a Forward Euler step of the inverse evolution gives a pair that satisfies the Backward Euler scheme of the forward equation; this is an algebraic identity, not a fit or a prediction. The high-order schemes (8)-(13) are Taylor expansions with the same reversal logic, so their consistency with high-order implicit schemes follows by construction from the definition of the reversed pairs, rather than from any fitted parameter. Accuracy of the generated pairs is checked against independently computed reference solutions in Table 1 and Figures 3-4, and downstream utility is measured with FNO and UNet on standard benchmarks (Tables 3-4). The only self-citation is reference [20], which is cited for introducing the phrase 'inverse evolution' and for its earlier use as a regularizer; the current paper re-derives the schemes and does not rely on [20] for any load-bearing theorem or uniqueness claim. Therefore no step in the derivation reduces to its inputs by definition. The main empirical weakness is that Section 3.3 compares N original pairs against 2N pairs (N original plus N generated), so the reported gains are compatible with the confound of doubled training-data quantity; this is a validity/control concern, not a circularity under the review criteria. Likewise, the initialization heuristic in Eq. (7) is unproved but is evaluated against external reference solutions rather than assumed circularly. Overall, the paper is largely self-contained, with only a minor, non-load-bearing self-citation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method relies on several unproved or underspecified choices: constants in Eq. (7) and Eq. (14), the validity of Taylor expansions for F and its derivatives, and the assumption that random combinations of snapshots remain near the solution manifold. These are method hyperparameters and domain assumptions, not fitted parameters, but they are load-bearing.

free parameters (4)

lambda_j = not reported
Weights in Eq. (7) for combining k randomly permuted time series to form inverse-evolution initializations. No values or ranges are given, and the quality of generated data depends on them.
C = random in [-0.1, 0.1] for Burgers with nu=0.001; otherwise not reported
Constant offset in Eqs. (7) and (14). The authors state that pseudo-spectral discretization permits large C for Navier-Stokes, but no criterion for choosing it is given.
a = 1 for Burgers nu=0.001; unspecified otherwise
Rescaling coefficient for preprocessing sharp-interface initializations in Eq. (14). The paper gives only that a lies in [0,1].
Delta t (inverse evolution step) = 0.05 for Burgers, 0.5 for Allen-Cahn and Navier-Stokes
The time interval of generated pairs, chosen to match the operator target t0. Table 1 shows accuracy and stability vary strongly with Delta t, so this choice is load-bearing.

assumptions (5)

domain assumption F is sufficiently differentiable so that the Taylor expansions (8)-(10) with derivatives (11)-(13) are valid.
Used in Section 2, 'High-order Schemes for Inverse Evolution'. This fails for solutions with shocks or sharp interfaces unless preprocessing is applied.
ad hoc to paper Reversed high-order Taylor inverse schemes correspond to stable high-order implicit schemes of the forward equation.
Stated after Eq. (13) without proof. The first-order case is exactly Backward Euler, but for second and third order the reversed expansion must be verified.
domain assumption Random combinations of original snapshots from Eq. (7) produce initial conditions that lie close enough to the solution manifold for inverse evolution to yield physically meaningful pairs.
Section 2, 'Initialization for Inverse Evolution'. The paper asserts nonlinearity ensures distinctness but gives no quantitative support.
standard math Periodic boundary conditions and regular rectangular domains are sufficient for the target applications.
Equation (1) and all experiments use periodic boundary conditions. The paper claims extension without loss of generality but does not test other boundary conditions.
domain assumption The 'true solutions' used for accuracy evaluation are trustworthy references.
Table 1 and Figures 3 and 4 compare generated data to reference solutions from PDEBench or standard solvers, with no independent verification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Inverse Evolution Data Augmentation for Neural PDE Solvers." pith.science (2026). https://pith.science/paper/F2ZDS4OK

@misc{pith2026250114604,
  author       = {Pith},
  title        = {Pith review of: Inverse Evolution Data Augmentation for Neural PDE Solvers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F2ZDS4OK}},
  note         = {Machine review of arXiv:2501.14604}
}

read the original abstract

Neural networks have emerged as promising tools for solving partial differential equations (PDEs), particularly through the application of neural operators. Training neural operators typically requires a large amount of training data to ensure accuracy and generalization. In this paper, we propose a novel data augmentation method specifically designed for training neural operators on evolution equations. Our approach utilizes insights from inverse processes of these equations to efficiently generate data from random initialization that are combined with original data. To further enhance the accuracy of the augmented data, we introduce high-order inverse evolution schemes. These schemes consist of only a few explicit computation steps, yet the resulting data pairs can be proven to satisfy the corresponding implicit numerical schemes. In contrast to traditional PDE solvers that require small time steps or implicit schemes to guarantee accuracy, our data augmentation method employs explicit schemes with relatively large time steps, thereby significantly reducing computational costs. Accuracy and efficacy experiments confirm the effectiveness of our approach. Additionally, we validate our approach through experiments with the Fourier Neural Operator and UNet on three common evolution equations that are Burgers' equation, the Allen-Cahn equation and the Navier-Stokes equation. The results demonstrate a significant improvement in the performance and robustness of the Fourier Neural Operator when coupled with our inverse evolution data augmentation method.

Figures

Figures reproduced from arXiv: 2501.14604 by the authors.

**Figure 1.** Preprocessing for Data with Sharp Interfaces We observe that solutions with poor differentiability can exacerbate the instability of the inverse evolution. For instance, solutions of Burgers’ equation often exhibit sharp interfaces, particularly when the viscosity approaches zero. During such instances, the solution may contain abrupt, nearly discontinuous changes, 8 [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗

**Figure 1.** Comparison between inverse evolution and forward evolution on [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗

**Figure 2.** Inverse evolution between original data and preprocessed data. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: Generated data for Allen-Cahn equation with [PITH_FULL_IMAGE:figures/full_fig_p011_3.png]

**Figure 4.** Figure 4: Generated data for Navier-Stokes equation with [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

**Figure 5.** Figure 5: Predictions of FNO trained on different datasets for Allen-Cahn [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 24 canonical work pages

[20]

Inverse Evolution Layers: Physics-informed Regularizers for Deep Neural Networks

Chaoyu Liu, Zhonghua Qiao, Chao Li, and Carola-Bibiane Sch¨ onlieb. Inverse evolution layers: Physics-informed regularizers for deep neural networks. arXiv preprint arXiv:2307.07344 , 2023

work page Pith review arXiv 2023
[1]

Representation equiv- alent neural operators: a framework for alias-free operator learning

Francesca Bartolucci, Emmanuel de Bezenac, Bogdan Raonic, Roberto Molinaro, Siddhartha Mishra, and Rima Alaifari. Representation equiv- alent neural operators: a framework for alias-free operator learning. Advances in Neural Information Processing Systems , 36, 2024

work page 2024
[2]

An introduction to fluid dynamics

George Keith Batchelor. An introduction to fluid dynamics . Cambridge university press, 1967. 17

work page 1967
[3]

Lie point symmetry data augmentation for neural pde solvers

Johannes Brandstetter, Max Welling, and Daniel E Worrall. Lie point symmetry data augmentation for neural pde solvers. In International Conference on Machine Learning , pages 2241–2256. PMLR, 2022

work page 2022
[4]

On the partial difference equations of mathematical physics

Richard Courant, Kurt Friedrichs, and Hans Lewy. On the partial difference equations of mathematical physics. IBM journal of Research and Development , 11(2):215–234, 1967

work page 1967
[5]

Scientific machine learning through physics–informed neural networks: Where we are and what’s next

Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next. Journal of Scientific Computing , 92(3):88, 2022

2022
[6]

Computational differential equations

Kenneth Eriksson. Computational differential equations . Cambridge University Press, 1996

work page 1996
[7]

Partial differential equations , volume 19

Lawrence C Evans. Partial differential equations , volume 19. American Mathematical Society, 2022

work page 2022

Show all 30 references

[8]

General covariance data augmentation for neural pde solvers

Vladimir Fanaskov, Tianchi Yu, Alexander Rudikov, and Ivan Oseledets. General covariance data augmentation for neural pde solvers. In In- ternational Conference on Machine Learning , pages 9665–9688. PMLR, 2023

2023
[9]

Towards multi- spatiotemporal-scale generalized pde modeling

Jayesh K Gupta and Johannes Brandstetter. Towards multi- spatiotemporal-scale generalized pde modeling. arXiv preprint arXiv:2209.15616, 2022

2022 arXiv
[10]

The era5 global reanalysis

Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, Andr´ as Hor´ anyi, Joaqu ´ ın Mu˜ noz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society , 146(730):1999–2049, 2020

1999
[11]

Adaptive moving mesh methods , volume 174

Weizhang Huang and Robert D Russell. Adaptive moving mesh methods , volume 174. Springer Science & Business Media, 2010

2010
[12]

On univer- sal approximation and error bounds for fourier neural operators

Nikola Kovachki, Samuel Lanthaler, and Siddhartha Mishra. On univer- sal approximation and error bounds for fourier neural operators. Journal of Machine Learning Research, 22(290):1–76, 2021

2021
[13]

Neural operator: Learning maps between function spaces with applications to pdes

Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research , 24(89):1–97, 2023

2023
[14]

Partial differential equations and stochastic methods in molecular dynamics

Tony Lelievre and Gabriel Stoltz. Partial differential equations and stochastic methods in molecular dynamics. Acta Numerica, 25:681–880, 2016. 18

2016
[15]

Finite difference methods for ordinary and partial differential equations: steady-state and time-dependent problems

Randall J LeVeque. Finite difference methods for ordinary and partial differential equations: steady-state and time-dependent problems . SIAM, 2007

2007
[16]

Physics-guided data augmentation for learning the solution operator of linear differential equations

Ye Li, Yiwen Pang, and Bin Shan. Physics-guided data augmentation for learning the solution operator of linear differential equations. In 2022 IEEE 8th International Conference on Cloud Computing and Intelligent Systems (CCIS) , pages 543–547. IEEE, 2022

2022
[17]

Fourier neural operator for parametric partial differential equations

Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. International Conference on Learning Representations , 2021

2021
[18]

Learning chaotic dynamics in dissipative systems

Zongyi Li, Miguel Liu-Schiaffini, Nikola Kovachki, Kamyar Azizzade- nesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Learning chaotic dynamics in dissipative systems. Advances in Neural Information Processing Systems , 35:16768–16781, 2022

2022
[19]

Physics-informed neural operator for learning partial differential equations

Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandku- mar. Physics-informed neural operator for learning partial differential equations. ACM/JMS Journal of Data Science , 2021

2021
[21]

Learning nonlinear operators via deeponet based on the universal approximation theorem of operators

Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelli- gence, 3(3):218–229, 2021

2021
[22]

Deep- xde: A deep learning library for solving differential equations

Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. Deep- xde: A deep learning library for solving differential equations. SIAM review, 63(1):208–228, 2021

2021
[23]

Numerical models for differential problems, volume 2

Alfio Quarteroni and Silvia Quarteroni. Numerical models for differential problems, volume 2. Springer, 2009

2009
[24]

Numerical approximation of partial differential equations, volume 23

Alfio Quarteroni and Alberto Valli. Numerical approximation of partial differential equations, volume 23. Springer Science & Business Media, 2008

2008
[25]

Physics- informed neural networks: A deep learning framework for solving forward 19 and inverse problems involving nonlinear partial differential equations

Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics- informed neural networks: A deep learning framework for solving forward 19 and inverse problems involving nonlinear partial differential equations. Journal of Computational physics , 378:686–707, 2019

2019
[26]

Numerical solution of partial differential equations: finite difference methods

Gordon D Smith. Numerical solution of partial differential equations: finite difference methods. Oxford university press, 1985

1985
[27]

Pdebench: An extensive benchmark for scientific machine learning

Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKin- lay, Francesco Alesiani, Dirk Pfl¨ uger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning. Advances in Neural Information Processing Systems , 35:1596–1611, 2022

2022
[28]

Long-time integration of paramet- ric evolution equations with physics-informed deeponets

Sifan Wang and Paris Perdikaris. Long-time integration of paramet- ric evolution equations with physics-informed deeponets. Journal of Computational Physics , 475:111855, 2023

2023
[29]

Learning the solu- tion operator of parametric partial differential equations with physics- informed deeponets

Sifan Wang, Hanwen Wang, and Paris Perdikaris. Learning the solu- tion operator of parametric partial differential equations with physics- informed deeponets. Science advances, 7(40):eabi8605, 2021

2021
[30]

Physics-informed machine learning for reliability and systems safety applications: State of the art and challenges

Yanwen Xu, Sara Kohtz, Jessica Boakye, Paolo Gardoni, and Pingfeng Wang. Physics-informed machine learning for reliability and systems safety applications: State of the art and challenges. Reliability Engi- neering & System Safety , 230:108900, 2023. 20

2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [20]

Inverse Evolution Layers: Physics-informed Regularizers for Deep Neural Networks

Chaoyu Liu, Zhonghua Qiao, Chao Li, and Carola-Bibiane Sch¨ onlieb. Inverse evolution layers: Physics-informed regularizers for deep neural networks. arXiv preprint arXiv:2307.07344 , 2023

work page Pith review arXiv 2023

[2] [1]

Representation equiv- alent neural operators: a framework for alias-free operator learning

Francesca Bartolucci, Emmanuel de Bezenac, Bogdan Raonic, Roberto Molinaro, Siddhartha Mishra, and Rima Alaifari. Representation equiv- alent neural operators: a framework for alias-free operator learning. Advances in Neural Information Processing Systems , 36, 2024

work page 2024

[3] [2]

An introduction to fluid dynamics

George Keith Batchelor. An introduction to fluid dynamics . Cambridge university press, 1967. 17

work page 1967

[4] [3]

Lie point symmetry data augmentation for neural pde solvers

Johannes Brandstetter, Max Welling, and Daniel E Worrall. Lie point symmetry data augmentation for neural pde solvers. In International Conference on Machine Learning , pages 2241–2256. PMLR, 2022

work page 2022

[5] [4]

On the partial difference equations of mathematical physics

Richard Courant, Kurt Friedrichs, and Hans Lewy. On the partial difference equations of mathematical physics. IBM journal of Research and Development , 11(2):215–234, 1967

work page 1967

[6] [5]

Scientific machine learning through physics–informed neural networks: Where we are and what’s next

Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next. Journal of Scientific Computing , 92(3):88, 2022

2022

[7] [6]

Computational differential equations

Kenneth Eriksson. Computational differential equations . Cambridge University Press, 1996

work page 1996

[8] [7]

Partial differential equations , volume 19

Lawrence C Evans. Partial differential equations , volume 19. American Mathematical Society, 2022

work page 2022