pith. machine review for the scientific record. sign in

arxiv: 2605.09523 · v2 · submitted 2026-05-10 · 💻 cs.LG · cs.CE· cs.NA· math.NA· physics.comp-ph· stat.ML

Recognition: no theorem link

HS-FNO: History-Space Fourier Neural Operator for Non-Markovian Partial Differential Equations

Lennon J. Shikhman
Authors on Pith no claims yet

Pith reviewed 2026-05-13 07:04 UTC · model grok-4.3

classification 💻 cs.LG cs.CEcs.NAmath.NAphysics.comp-phstat.ML
keywords neural operatorsnon-Markovian PDEsdelay equationshistory spaceFourier neural operatorautoregressive predictionsurrogate modeling
0
0 comments X

The pith

HS-FNO halves rollout error for non-Markovian PDEs by lifting states to include exact history shifts instead of learning them.

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

Standard neural operators assume the instantaneous field is a complete state, but this breaks for delay equations and memory-driven systems where identical present states can lead to different futures. HS-FNO lifts the state to a history window u_t(θ,x) with θ in [-τ,0] and decomposes each update into a learned predictor for the newly exposed future slice plus an exact shift-append transport for the known portion of the window. This enforces the natural discrete history evolution, cuts the learned output dimension, and yields lower one-step, history-space, and rollout errors than current-state, lag-stack, or unconstrained history-to-history baselines across five benchmark families. The gains are largest in autoregressive rollouts, where aggregate error drops from 0.241-0.188 to 0.094 while using fewer parameters.

Core claim

HS-FNO formulates a neural operator directly on the lifted history-state field and splits each time step into a learned future-slice predictor and an exact shift-append transport; this structure produces the lowest aggregate errors on delayed reaction-diffusion, spatial epidemiology, nonlocal neural fields, delayed waves, and distributed-memory closures, with the clearest improvement appearing in long autoregressive prediction.

What carries the argument

History-state lifting to u_t(θ,x) combined with learned future-slice prediction plus exact shift-append transport.

If this is right

  • Autoregressive forecasts of delay and nonlocal PDEs become stable enough for practical surrogate use without retraining at each step.
  • The same model size produces lower error than unconstrained history-to-history operators, freeing parameters for finer spatial resolution.
  • The inductive bias applies across reaction-diffusion, epidemiology, neural-field, and wave benchmarks without per-family redesign.
  • One-step and history-space errors also improve, indicating the structure helps even when full rollouts are not required.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same split-predictor-plus-exact-transport pattern could be inserted into other operator families beyond Fourier bases to handle memory effects.
  • Testing on systems with slowly decaying memory kernels would reveal whether the fixed-tau window must be made adaptive.
  • If the shift-append step is replaced by a learned transport with small regularization, error accumulation might be further reduced on very long horizons.

Load-bearing premise

A fixed finite history window of length tau is sufficient to capture all relevant non-Markovian memory effects and the shift-append step stays numerically stable over long rollouts.

What would settle it

Measure whether rollout error stays below 0.12 when the same trained model is tested on trajectories whose required memory exceeds the fixed tau or when run for 10x longer autoregressive steps than the training horizon.

Figures

Figures reproduced from arXiv: 2605.09523 by Lennon J. Shikhman.

Figure 1
Figure 1. Figure 1: HS-FNO architecture. The predictor PΘ receives the current history field ut(θ, x) together with conditioning variables (µ, τ, ∆t) and predicts the newly exposed future slice ub(t + ∆t, ·). The known portion of the history is transported exactly, and ShiftAppend combines this with the predicted slice to form the updated history GΘ(ut). Shift-append history evolution. For 0 < ∆t ≤ τ , the exact update satisf… view at source ↗
Figure 1
Figure 1. Figure 1: HS-FNO architecture. The predictor PΘ receives the current history field ut(θ, x) together with conditioning variables (µ, τ, ∆t) and predicts the newly exposed future slice ub(t + ∆t, ·). The known portion of the history is transported exactly, and ShiftAppend combines this with the predicted slice to form the updated history GΘ(ut). Shift-append history evolution For 0 < ∆t ≤ τ , the exact update satisfi… view at source ↗
Figure 2
Figure 2. Figure 2: Aggregate relative errors across benchmark–regime cells. Bars show ten-seed [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Aggregate relative errors across benchmark–regime cells. Bars show ten-seed means and error bars show 95% [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean rollout error over autoregressive prediction steps. HS-FNO has the lowest [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean rollout error over autoregressive prediction steps. HS-FNO has the lowest error at every rollout step [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Real-world traffic sanity check under the standard 12-input/12-output protocol. [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Real-world traffic sanity check under the standard 12-input/12-output protocol. Bars show denormalized [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
read the original abstract

Neural operators provide fast surrogate models for time-dependent partial differential equations, but their standard autoregressive use usually assumes that the instantaneous field $u(t,\cdot)$ is a complete state. This assumption fails for delay equations, distributed-memory systems, and other non-Markovian dynamics: two trajectories may agree at time $t$ and nevertheless have different futures because their histories differ. We introduce the History-Space Fourier Neural Operator (HS-FNO), a neural operator for delay and memory-driven PDEs formulated on the lifted state $u_t(\theta,x)=u(t+\theta,x)$, $\theta\in[-\tau,0]$. The key computational step is to decompose one history-state update into a learned predictor for the newly exposed future slice and an exact shift-append transport for the portion of the history window already known from the previous state. This avoids learning deterministic history coordinates, reduces the learned output dimension, and enforces the natural discrete history update. We test HS-FNO on five benchmark families covering delayed reaction--diffusion, spatial epidemiology, nonlocal neural-field dynamics, delayed waves, and distributed-memory closures. Across ten random seeds, HS-FNO attains the lowest aggregate one-step, history-space, and rollout errors among the principal baselines. The largest gain occurs in autoregressive prediction, where aggregate rollout error decreases from $0.241$, $0.188$, and $0.185$ for current-state, lag-stack, and unconstrained history-to-history operators, respectively, to $0.094$. The same model uses fewer parameters than unconstrained history prediction. These results indicate that enforcing the discrete shift structure of history-state evolution is an effective inductive bias for non-Markovian PDE surrogate modeling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces the History-Space Fourier Neural Operator (HS-FNO) for non-Markovian time-dependent PDEs. It lifts the state to a history window u_t(θ, x) with θ ∈ [-τ, 0] and decomposes each update into a learned neural-operator prediction for the newly exposed future slice plus an exact shift-append transport on the known portion of the history. Experiments across five benchmark families (delayed reaction-diffusion, spatial epidemiology, nonlocal neural fields, delayed waves, distributed-memory closures) and ten random seeds show HS-FNO attaining the lowest aggregate one-step, history-space, and rollout errors, with rollout error dropping from 0.185 (unconstrained history-to-history) to 0.094 while using fewer parameters.

Significance. If the empirical gains hold under closer scrutiny, the explicit separation of learned prediction from exact discrete transport supplies a useful inductive bias for surrogate modeling of memory-driven systems, improving long-horizon autoregressive accuracy and parameter efficiency. The architecture's reduction of the learned output dimension via the exact-transport step is a concrete strength that could transfer to other operator families.

major comments (3)
  1. [Abstract and Experiments] Abstract and Experiments: the central claim that HS-FNO attains the lowest aggregate rollout error (0.094 versus 0.185 for the unconstrained baseline) is reported without details on exact baseline implementations, hyperparameter-matching protocols, or statistical significance testing of the gap across the ten seeds. This information is load-bearing for the empirical superiority statement.
  2. [Method] Method (history-state update decomposition): the shift-append transport is exact only when the incoming history is perfect. Once the learned future-slice predictor introduces approximation error, that error is shifted into subsequent windows; the manuscript provides no analysis or per-benchmark error-growth curves showing how such drift behaves over rollouts longer than τ, which directly affects the autoregressive-prediction claim.
  3. [Experiments] Experiments: the formulation assumes a fixed finite τ suffices to capture all relevant non-Markovian effects for each benchmark family, yet no τ-sensitivity sweeps or comparison of rollout horizons against the intrinsic memory scale of the PDEs are presented. This limits evaluation of whether the reported gains generalize beyond the chosen window lengths.
minor comments (2)
  1. [Method] Notation: the lifted state u_t(θ, x) and the precise mechanics of the shift-append operation would benefit from an explicit equation or schematic diagram in the method section.
  2. [Experiments] References: the principal baselines (current-state, lag-stack, unconstrained history-to-history) should be cited with their original papers to allow readers to verify implementation details.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below with clarifications from the manuscript and outline specific revisions that will be incorporated to strengthen the empirical claims and analysis.

read point-by-point responses

  1. Referee: [Abstract and Experiments] Abstract and Experiments: the central claim that HS-FNO attains the lowest aggregate rollout error (0.094 versus 0.185 for the unconstrained baseline) is reported without details on exact baseline implementations, hyperparameter-matching protocols, or statistical significance testing of the gap across the ten seeds. This information is load-bearing for the empirical superiority statement.

    Authors: We agree that the current presentation lacks sufficient detail on baseline implementations and statistical validation. The manuscript already specifies the three baselines (current-state FNO, lag-stack FNO, unconstrained history-to-history FNO) and reports aggregate errors over ten seeds, but does not describe the hyperparameter search or significance tests. In the revised version we will add a dedicated experimental-setup subsection that (i) gives the exact architecture and training protocol for each baseline, (ii) documents the grid-search ranges used to match model capacity and training effort, and (iii) reports mean, standard deviation, and paired t-test p-values for the rollout-error differences across the ten seeds. These additions will directly support the superiority statement. revision: yes

  2. Referee: [Method] Method (history-state update decomposition): the shift-append transport is exact only when the incoming history is perfect. Once the learned future-slice predictor introduces approximation error, that error is shifted into subsequent windows; the manuscript provides no analysis or per-benchmark error-growth curves showing how such drift behaves over rollouts longer than τ, which directly affects the autoregressive-prediction claim.

    Authors: The referee correctly notes that approximation error introduced by the learned slice predictor will be exactly transported forward by the shift-append step. While the decomposition guarantees that known history coordinates are never re-learned, the manuscript indeed omits explicit long-horizon drift analysis. We will therefore add, in the revised Experiments section, per-benchmark error-growth curves for autoregressive rollouts extending to at least 5τ–10τ. These curves will quantify the accumulation of drift for HS-FNO versus the baselines and will be accompanied by a short discussion of how the exact-transport step limits error growth relative to fully learned history-to-history mappings. revision: yes

  3. Referee: [Experiments] Experiments: the formulation assumes a fixed finite τ suffices to capture all relevant non-Markovian effects for each benchmark family, yet no τ-sensitivity sweeps or comparison of rollout horizons against the intrinsic memory scale of the PDEs are presented. This limits evaluation of whether the reported gains generalize beyond the chosen window lengths.

    Authors: For each benchmark the value of τ was chosen to match the explicit delay or memory scale stated in the PDE definition (e.g., the fixed delay in the delayed reaction-diffusion and wave equations). Nevertheless, the manuscript does not present sensitivity sweeps. In the revision we will add τ-sensitivity plots for two representative families (delayed reaction-diffusion and nonlocal neural fields), varying τ around the nominal value and reporting rollout error versus τ. We will also include a brief discussion, referencing the benchmark descriptions in Section 4, that relates the chosen τ to the intrinsic memory scale of each PDE family. These results will be placed in the main Experiments section or as supplementary material. revision: partial

Circularity Check

0 steps flagged

No circularity: architectural design with exact transport and external benchmark validation

full rationale

The paper defines HS-FNO via an explicit decomposition of each history-state update into a learned predictor for the new future slice plus an exact shift-append transport on the known history window. This is presented as an inductive bias that reduces output dimension and enforces the discrete update rule, without any derivation that equates the claimed performance gains to fitted parameters or prior self-citations. All reported results (one-step, history-space, and rollout errors across five benchmark families and ten seeds) are measured against independent baselines on external test trajectories; they do not reduce by construction to quantities defined inside the model equations. No uniqueness theorems, ansatzes, or self-citations are invoked to justify the central claims.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The method rests on the domain assumption that a fixed finite history window suffices and on standard neural-network training assumptions; no new physical entities are postulated.

free parameters (1)
  • neural network weights and biases
    Standard trainable parameters of the Fourier layers and predictor network; fitted during training on the benchmark data.
axioms (2)
  • domain assumption A fixed finite history window of length tau captures all relevant memory effects for the target PDE families.
    Invoked when the state is lifted to u_t(theta,x) for theta in [-tau,0] and when the shift-append operation is treated as exact.
  • standard math Fourier Neural Operator layers can be applied to the history-augmented field without loss of the underlying operator-learning guarantees.
    Assumed when the architecture re-uses the standard FNO backbone on the lifted state.
invented entities (1)
  • History-space state u_t(theta,x) no independent evidence
    purpose: To represent the non-Markovian lifted state that includes the full history window.
    New state representation introduced to enable the exact shift-append transport.

pith-pipeline@v0.9.0 · 5625 in / 1715 out tokens · 54245 ms · 2026-05-13T07:04:24.014568+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages

  1. [1]

    Hale and Sjoerd M

    Jack K. Hale and Sjoerd M. Verduyn Lunel.Introduction to Functional Differential Equations, volume 99 of Applied Mathematical Sciences. Springer, 1993

    work page 1993

  2. [2]

    Springer, 1996

    Jianhong Wu.Theory and Applications of Partial Functional Differential Equations, volume 119 ofApplied Mathematical Sciences. Springer, 1996

    work page 1996

  3. [3]

    Threshold dynamics of a delayed reaction diffusion equation subject to the dirichlet condition.Journal of Biological Dynamics, 3(2-3):331–341, 2009

    Taishan Yi, Yuming Chen, and Jianhong Wu. Threshold dynamics of a delayed reaction diffusion equation subject to the dirichlet condition.Journal of Biological Dynamics, 3(2-3):331–341, 2009. doi: 10.1080/17513750802425

    work page doi:10.1080/17513750802425 2009

  4. [4]

    PMID: 22880838

    URL https://doi.org/10.1080/17513750802425656. PMID: 22880838

    work page doi:10.1080/17513750802425656

  5. [5]

    Marc R. Roussel. The use of delay differential equations in chemical kinetics.The Journal of Physical Chemistry, 100(20):8323–8330, 1996. doi: 10.1021/jp9600672

    work page doi:10.1021/jp9600672 1996

  6. [6]

    S. M. Oliva. Reaction–diffusion equations with nonlinear boundary delay.Journal of Dynamics and Differential Equations, 11:279–296, 1999. doi: 10.1023/A:1021929413376

    work page doi:10.1023/a:1021929413376 1999

  7. [7]

    Parish and Karthik Duraisamy

    Eric J. Parish and Karthik Duraisamy. Non-markovian closure models for large eddy simulations using the mori-zwanzig formalism.Phys. Rev. Fluids, 2:014604, Jan 2017. doi: 10.1103/PhysRevFluids.2.014604. URL https://link.aps.org/doi/10.1103/PhysRevFluids.2.014604. 12 HS-FNO for Non-Markovian PDEsPREPRINT

    work page doi:10.1103/physrevfluids.2.014604 2017

  8. [8]

    Parish and Karthik Duraisamy

    Eric J. Parish and Karthik Duraisamy. A dynamic subgrid scale model for large eddy simulations based on the mori–zwanzig formalism.Journal of Computational Physics, 349:154–175, 2017. ISSN 0021-9991. doi: https: //doi.org/10.1016/j.jcp.2017.07.053. URL https://www.sciencedirect.com/science/article/pii/S0021999117305612

    work page doi:10.1016/j.jcp.2017.07.053 2017

  9. [9]

    Sorokin and Andrei V

    Vsevolod G. Sorokin and Andrei V . Vyazmin. Nonlinear reaction–diffusion equations with delay: Partial survey, exact solutions, test problems, and numerical integration.Mathematics, 10(11), 2022. ISSN 2227-7390. doi: 10.3390/math10111886. URL https://www.mdpi.com/2227-7390/10/11/1886

    work page doi:10.3390/math10111886 2022

  10. [10]

    C.V . Pao. Finite difference solutions of reaction diffusion equations with continuous time delays.Computers & Mathematics with Applications, 42(3):399–412, 2001. ISSN 0898-1221. doi: https://doi.org/10.1016/S0898-122 1(01)00165-1. URL https://www.sciencedirect.com/science/article/pii/S0898122101001651

    work page doi:10.1016/s0898-122 2001

  11. [11]

    Yuan-Ming Wang. Asymptotic behavior of the numerical solutions of time-delayed reaction diffusion equations with non-monotone reaction term.ESAIM: Mathematical Modelling and Numerical Analysis, 37(2):259–276,

    work page

  12. [12]

    doi: 10.1051/m2an:2003025

    work page doi:10.1051/m2an:2003025

  13. [13]

    Yuan-Ming Wang and C. V . Pao. Time-delayed finite difference reaction-diffusion systems with nonquasimonotone functions.Numerische Mathematik, 103:485–513, 2006. doi: 10.1007/s00211-006-0685-y

    work page doi:10.1007/s00211-006-0685-y 2006

  14. [14]

    Amirali, G

    I. Amirali, G. M. Amiraliyev, M. Cakir, and E. Cimen. Explicit finite difference methods for the delay pseu- doparabolic equations.The Scientific World Journal, 2014:497393, 2014. doi: 10.1155/2014/497393

    work page doi:10.1155/2014/497393 2014

  15. [15]

    Galerkin finite element method for the generalized delay reaction- diffusion equation.Research in Mathematics, 9:1–16, 12 2022

    Gemeda Lubo and Gemechis File Duressa. Galerkin finite element method for the generalized delay reaction- diffusion equation.Research in Mathematics, 9:1–16, 12 2022. doi: 10.1080/27684830.2022.2071388

    work page doi:10.1080/27684830.2022.2071388 2022

  16. [16]

    Yao-Lung L. Fang. Fdtd: Solving 1+1d delay pde in parallel.Computer Physics Communications, 235:422–432,

    work page

  17. [17]

    doi: https://doi.org/10.1016/j.cpc.2018.08.018

    ISSN 0010-4655. doi: https://doi.org/10.1016/j.cpc.2018.08.018. URL https://www.sciencedirect.com/scie nce/article/pii/S001046551830314X

    work page doi:10.1016/j.cpc.2018.08.018 2018

  18. [18]

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

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378:686–707, 2019. ISSN 0021-9991. doi: 10.1016/j.jcp.2018.10.045. URL https://www.sciencedirect. com/science/article/pii/...

    work page doi:10.1016/j.jcp.2018.10.045 2019

  19. [19]

    Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

    George Karniadakis, Yannis Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3:422–440, 2021. doi: 10.1038/s42254-021-00314-5

    work page doi:10.1038/s42254-021-00314-5 2021

  20. [20]

    Double-activation neural network for solving parabolic equations with time delay

    Qiumei Huang and Qiao Zhu. Double-activation neural network for solving parabolic equations with time delay. Neurocomputing, 635:129978, 2025. ISSN 0925-2312. doi: https://doi.org/10.1016/j.neucom.2025.129978. URL https://www.sciencedirect.com/science/article/pii/S0925231225006502

    work page doi:10.1016/j.neucom.2025.129978 2025

  21. [21]

    A deep neural network framework for solving forward and inverse problems in delay differential equations.Journal of Computational and Applied Mathematics, 477:117154, 2026

    Housen Wang, Yuxing Chen, Sirong Cao, Xiaoli Wang, and Qiang Liu. A deep neural network framework for solving forward and inverse problems in delay differential equations.Journal of Computational and Applied Mathematics, 477:117154, 2026. ISSN 0377-0427. doi: https://doi.org/10.1016/j.cam.2025.117154. URL https://www.sciencedirect.com/science/article/pii/...

    work page doi:10.1016/j.cam.2025.117154 2026

  22. [22]

    Modeling of high- dimensional time-delay chaotic system based on fourier neural operator.Chaos, Solitons & Fractals, 188:115523,

    Jiacheng Feng, Lin Jiang, Lianshan Yan, Xingchen He, Anlin Yi, Wei Pan, and Bin Luo. Modeling of high- dimensional time-delay chaotic system based on fourier neural operator.Chaos, Solitons & Fractals, 188:115523,

    work page

  23. [23]

    doi: https://doi.org/10.1016/j.chaos.2024.115523

    ISSN 0960-0779. doi: https://doi.org/10.1016/j.chaos.2024.115523. URL https://www.sciencedirect.com/ science/article/pii/S0960077924010750

    work page doi:10.1016/j.chaos.2024.115523 2024

  24. [24]

    Learning stochastic dynamics with statistics-informed neural network.Journal of Computational Physics, 474:111819, 2023

    Yuanran Zhu, Yu-Hang Tang, and Changho Kim. Learning stochastic dynamics with statistics-informed neural network.Journal of Computational Physics, 474:111819, 2023. ISSN 0021-9991. doi: 10.1016/j.jcp.2022.111819. URL https://www.sciencedirect.com/science/article/pii/S0021999122008828

    work page doi:10.1016/j.jcp.2022.111819 2023

  25. [25]

    Neural delay differential equations

    Qunxi Zhu, Yao Guo, and Wei Lin. Neural delay differential equations. InInternational Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=Q1jmmQz72M2

    work page 2021

  26. [26]

    Neural piecewise-constant delay differential equations

    Qunxi Zhu, Yifei Shen, Dongsheng Li, and Wei Lin. Neural piecewise-constant delay differential equations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 2022. doi: 10.1609/aaai.v36i8.20911

    work page doi:10.1609/aaai.v36i8.20911 2022

  27. [27]

    Learning stable deep dynamics models for partially observed or delayed dynamical systems

    Andreas Schlaginhaufen, Philippe Wenk, Andreas Krause, and Florian Dorfler. Learning stable deep dynamics models for partially observed or delayed dynamical systems. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, volume 34, pages 11870–11882. Curran Associates, Inc...

    work page 2021

  28. [28]

    Learning the delay in delay differential equations

    Robert Stephany, Maria Antonia Oprea, Gabriella Torres Nothaft, Mark Walth, Arnaldo Rodriguez-Gonzalez, and William A Clark. Learning the delay in delay differential equations. InICLR 2024 Workshop on AI4DifferentialEquations In Science, 2024. URL https://openreview.net/forum?id=VTYhJLoOaR. 13 HS-FNO for Non-Markovian PDEsPREPRINT

    work page 2024

  29. [29]

    Neural adaptive delay differential equations.Neurocomputing, 648:130634, 2025

    Chao Zhou, Qieshi Zhang, and Jun Cheng. Neural adaptive delay differential equations.Neurocomputing, 648:130634, 2025. ISSN 0925-2312. doi: https://doi.org/10.1016/j.neucom.2025.130634. URL https: //www.sciencedirect.com/science/article/pii/S0925231225013062

    work page doi:10.1016/j.neucom.2025.130634 2025

  30. [30]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3): 218–229, 2021

    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 Intelligence, 3(3): 218–229, 2021

    work page 2021

  31. [31]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. InInterna- tional Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=c8P9NQVtmnO

    work page 2021

  32. [32]

    Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

    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. URL http://jmlr.org/papers/v24/21-1524.html

    work page 2023

  33. [33]

    Geometry-informed neural operator for large-scale 3d pdes, 2023

    Zongyi Li, Nikola Borislavov Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Prakash Otta, Moham- mad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, and Anima Anandkumar. Geometry-informed neural operator for large-scale 3d pdes, 2023. URL https://arxiv.org/abs/2309.00583

    work page arXiv 2023

  34. [34]

    Convolutional neural operators for robust and accurate learning of pdes

    Bogdan Raoni´c, Roberto Molinaro, Tim De Ryck, Tobias Rohner, Francesca Bartolucci, Rima Alaifari, Siddhartha Mishra, and Emmanuel de Bézenac. Convolutional neural operators for robust and accurate learning of pdes. In Advances in Neural Information Processing Systems, volume 36, pages 77187–77200, 2023. URL https://proceedi ngs.neurips.cc/paper_files/pap...

    work page 2023

  35. [35]

    U-no: U-shaped neural operators.arXiv preprint arXiv:2204.11127, 2022

    Md Ashiqur Rahman, Zachary E. Ross, and Kamyar Azizzadenesheli. U-no: U-shaped neural operators, 2023. URL https://arxiv.org/abs/2204.11127

    work page arXiv 2023

  36. [36]

    Wang, Yuan Yin, Jean-Noël Vittaut, and Patrick Gallinari

    Louis Serrano, Lise Le Boudec, Armand Kassaï Koupaï, Thomas X. Wang, Yuan Yin, Jean-Noël Vittaut, and Patrick Gallinari. Operator learning with neural fields: Tackling pdes on general geometries. InAdvances in Neural Information Processing Systems, volume 36, pages 70581–70611, 2023. URL https://proceedings.neurips. cc/paper_files/paper/2023/file/df543023...

    work page 2023

  37. [37]

    Choose a transformer: Fourier or galerkin, 2021

    Shuhao Cao. Choose a transformer: Fourier or galerkin, 2021. URL https://arxiv.org/abs/2105.14995

    work page arXiv 2021

  38. [38]

    Transformer for partial differential equations’ operator learning.arXiv preprint arXiv:2205.13671, 2022

    Zijie Li, Kazem Meidani, and Amir Barati Farimani. Transformer for partial differential equations’ operator learning, 2023. URL https://arxiv.org/abs/2205.13671

    work page arXiv 2023

  39. [39]

    Latent neural operator for solving forward and inverse pde problems

    Tian Wang and Chuang Wang. Latent neural operator for solving forward and inverse pde problems. InAdvances in Neural Information Processing Systems, volume 37, pages 33085–33107, 2024. doi: 10.52202/079017-1042. URL https://proceedings.neurips.cc/paper_files/paper/2024/file/39f6d5c2e310a5a629dcfc4d517aa0d1-Paper-C onference.pdf

    work page doi:10.52202/079017-1042 2024

  40. [40]

    arXiv preprint arXiv:2111.03794 , year =

    Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations, 2023. URL https://arxiv.org/abs/2111.03794

    work page arXiv 2023

  41. [41]

    O’Leary-Roseberry, P

    Thomas O’Leary-Roseberry, Peng Chen, Umberto Villa, and Omar Ghattas. Derivative-informed neural operator: An efficient framework for high-dimensional parametric derivative learning.Journal of Computational Physics, 496:112555, 2024. ISSN 0021-9991. doi: 10.1016/j.jcp.2023.112555. URL https://www.sciencedirect.com/scienc e/article/pii/S0021999123006502

    work page doi:10.1016/j.jcp.2023.112555 2024

  42. [42]

    PDEBench: An extensive benchmark for scientific machine learning

    Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Dan MacKinlay, Francesco Alesiani, Dirk Pflüger, and Mathias Niepert. PDEBench: An extensive benchmark for scientific machine learning. InAdvances in Neural Information Processing Systems Datasets and Benchmarks Track, 2022. URL https://openreview.net/forum?id=dh _MkX0QfrK

    work page 2022

  43. [43]

    Detecting strange attractors in turbulence

    Floris Takens. Detecting strange attractors in turbulence. In David Rand and Lai-Sang Young, editors,Dynamical Systems and Turbulence, Warwick 1980, pages 366–381, Berlin, Heidelberg, 1981. Springer Berlin Heidelberg. ISBN 978-3-540-38945-3

    work page 1980

  44. [44]

    Yorke, and Martin Casdagli

    Tim Sauer, James A. Yorke, and Martin Casdagli. Embedology.Journal of Statistical Physics, 65:579–616, 1991. doi: 10.1007/BF01053745

    work page doi:10.1007/bf01053745 1991

  45. [45]

    Memory effects in irreversible thermodynamics.Phys

    Robert Zwanzig. Memory effects in irreversible thermodynamics.Phys. Rev., 124:983–992, Nov 1961. doi: 10.1103/PhysRev.124.983. URL https://link.aps.org/doi/10.1103/PhysRev.124.983

    work page doi:10.1103/physrev.124.983 1961

  46. [46]

    Chorin, Ole H

    Alexandre J. Chorin, Ole H. Hald, and Raz Kupferman. Optimal prediction and the mori–zwanzig representation of irreversible processes.Proceedings of the National Academy of Sciences, 97(7):2968–2973, 2000. doi: 10.1073/pnas.97.7.2968. URL https://www.pnas.org/doi/abs/10.1073/pnas.97.7.2968. 14 HS-FNO for Non-Markovian PDEsPREPRINT

    work page doi:10.1073/pnas.97.7.2968 2000

  47. [47]

    Incorporation of memory effects in coarse-grained modeling via the mori-zwanzig formalism.The Journal of Chemical Physics, 143(24):243128, 11 2015

    Zhen Li, Xin Bian, Xiantao Li, and George Em Karniadakis. Incorporation of memory effects in coarse-grained modeling via the mori-zwanzig formalism.The Journal of Chemical Physics, 143(24):243128, 11 2015. ISSN 0021-9606. doi: 10.1063/1.4935490. URL https://doi.org/10.1063/1.4935490

    work page doi:10.1063/1.4935490 2015

  48. [48]

    Parish, and Karthik Duraisamy

    Ayoub Gouasmi, Eric J. Parish, and Karthik Duraisamy. A priori estimation of memory effects in reduced-order models of nonlinear systems using the mori-zwanzig formalism.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 473(2205):20170385, 2017. doi: 10.1098/rspa.2017.0385

    work page doi:10.1098/rspa.2017.0385 2017

  49. [49]

    Long short-term memory.Neural Comput., 9(8):1735–1780, November

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory.Neural Comput., 9(8):1735–1780, November

    work page

  50. [50]

    Neural Computation 9, 1735–1780

    ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URL https://doi.org/10.1162/neco.1997.9.8.1735

    work page doi:10.1162/neco.1997.9.8.1735 1997

  51. [51]

    Convolutional lstm network: A machine learning approach for precipitation nowcasting

    Xingjian SHI, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun WOO. Convolutional lstm network: A machine learning approach for precipitation nowcasting. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. URL https://proceed...

    work page 2015

  52. [52]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL ht...

    work page 2017

  53. [53]

    Diffusion convolutional recurrent neural network: Data-driven traffic forecasting

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. InInternational Conference on Learning Representations (ICLR ’18), 2018. 15

    work page 2018