Pith. sign in

REVIEW 3 major objections 5 minor 63 references

Time Resolution Independent Operator Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read NCDE-DeepONet claims a single neural operator can learn transient PDE solutions from irregularly sampled load histories and answer queries at any space–time point without retraining.

desk verdict Solid incremental architecture with clean fixed-grid results, but the resolution-independence claims need much stronger evidence than the current experiments provide. read the letter →

arxiv 2507.02524 v1 pith:LCRRSTCH submitted 2025-07-03 cs.CE cs.NAmath.NA

classification cs.CEcs.NAmath.NA
keywords neuralcontrolleddifferentialequationsoperatorlearningDeepONettransientmechanicstimeresolutionindependenceirregularserieselastodynamicsthermoelasticity
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 introduces NCDE-DeepONet, a neural operator that learns the mapping from time-varying boundary conditions to full space–time solution fields of transient PDEs. The central claim is that a Neural Controlled Differential Equation in the branch encodes the entire load history as a continuous controlled path, so the model can accept inputs sampled at different rates without retraining. Because the trunk network takes explicit $(x,t)$ coordinates as input, outputs can be queried at arbitrary spatial locations and instants, including times and meshes never seen during training. On transient Poisson, elastodynamic, and thermoelastic benchmarks, the model reports mean relative $\mathrm{L}^2$ errors around $7\times10^{-3}$ and reduces the mean error on elastodynamics by roughly 40% compared with a GRU-DeepONet baseline. If correct, this provides a single resolution-independent surrogate for transient mechanics that can be interrogated at any point in space and time.

What carries the argument

The central mechanism is the Neural Controlled Differential Equation (NCDE), the continuous-time analogue of an RNN, in which a hidden state $z(t)$ evolves under $\mathrm{d}z(t) = f_\theta(z(t))\,\mathrm{d}X(t)$ with $X(t)$ the cubic Hermite interpolation of the observed load history. This controlled ODE makes the branch encoding insensitive to how densely or irregularly the input is sampled, because the interpolated path drives the dynamics continuously. The second half of the machinery is a DeepONet-style trunk network that maps the query point $(x,t)$ into the same latent space and returns the predicted field through the inner product $\langle b, t\rangle$ of branch and trunk outputs; this is what makes output queries valid at arbitrary times and locations. Training uses the adjoint sensitivity method, re-integrating the ODE backward to obtain gradients with constant memory. Together these two pieces decouple the input signal's discretization from the output query's discreteness.

What would settle it

Run the same NCDE-DeepONet on load histories whose observation times are drawn from a random, non-uniform process (for example, Poisson arrivals) with matched average sample counts, and compare mean relative $\mathrm{L}^2$ error to the regular-grid baseline; if the error rises substantially while the GRU baseline stays flat, the interpolation-adequacy premise is refuted, and if it holds, the resolution-independence claim is corroborated on genuinely irregular data.

Watch

Extended reading notes

Core claim

The paper's discovery is that making the branch of a sequential DeepONet a Neural Controlled Differential Equation, and supplying the trunk with space–time coordinates, removes the fixed-grid bottleneck that ties recurrent operator learners to their training discretization. The NCDE interpolates the discrete load samples into a continuous control path $X(t)$ and evolves a hidden state $z(t)$ according to $\mathrm{d}z(t) = f_\theta(z(t))\,\mathrm{d}X(t)$, so any change in sampling rate is absorbed by the interpolation path rather than by a discrete recurrence. The trunk, fed arbitrary $(x,t)$, returns an inner product with this latent encoding that yields the field value $u(x,t)$. On three benchmark PDEs the trained model predicts fields with mean relative $\mathrm{L}^2$ errors near $10^{-3}$, and a dedicated experiment shows consistent spatial fields when input heat fluxes are sampled at 50%, 100%, and 200% of the training density. The paper frames this as a single framework that is both input-resolution independent and output-resolution independent, closing a gap that spatial-resolution-only methods like RINO leave open.

Load-bearing premise

The claim of input-resolution independence rests on the assumption that cubic Hermite interpolation of the discrete observations reproduces the true load signal closely enough that the learned encoder, trained on evenly spaced samples, remains accurate at sampling rates it never saw; the paper tests only subsampling and upsampling of the same regular signals, not genuinely irregular gaps, missing chunks, or sensor noise.

Editorial extensions

If this is right

  • A single trained model can be queried at arbitrary output times and spatial locations, so downstream analyses can refine or re-mesh the solution without retraining or interpolating either the input or the output.
  • Deployment on sensor streams with changing sampling rates, dropouts, or upsampling needs becomes seamless, since the branch re-encodes any discretization of the same history at inference time.
  • The roughly 40% mean-error reduction over GRU-DeepONet on elastodynamics suggests continuous controlled paths are a better inductive bias than discrete recurrence for wave-like, path-dependent dynamics even on regularly sampled data.
  • Consistent predictions across 50%, 100%, and 200% input sampling on the thermoelastic benchmark indicate that the learned latent trajectory, not the underlying time grid, carries the solution information.
  • Because the same hyperparameters transfer across Poisson, elastodynamics, and coupled thermoelasticity, the architecture appears to be a problem-agnostic recipe for time-dependent operator learning.

Reading between the lines

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

  • A direct stress test the paper does not run: feed genuinely irregular arrival times (random gaps) rather than subsampled versions of a regular grid; if accuracy degrades sharply at large gaps, the practical scope of 'irregular sampling' is narrower than the abstract claims.
  • The same branch–trunk decoupling could transfer temporal resolution independence to spatial resolution independence by interpolating spatially scattered sensors into a control field, effectively merging NCDE-DeepONet's temporal idea with RINO's spatial idea.
  • The reported mean relative errors aggregate over all test samples; near load boundaries errors are visibly larger, so safety-critical use would need per-region error bounds or an uncertainty estimate before relying on instantaneous queries.
  • A head-to-head comparison with time-adaptive operators (TI-DeepONet, TANTE) on long-horizon extrapolation would clarify whether continuous control paths or explicit learned time-stepping is the more durable route to resolution-independent forecasting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes NCDE-DeepONet, an operator-learning architecture that replaces the recurrent branch of a sequential DeepONet with a Neural Controlled Differential Equation and augments the trunk with space–time coordinates. The central claim is that the NCDE branch makes the learned operator input-resolution-independent (handling irregularly sampled or resampled load histories without retraining) and that the spatiotemporal trunk makes it output-resolution-independent (permitting queries at arbitrary spatial locations and times). The authors report relative L2 errors around 7e-3 on transient Poisson, elastodynamics, and thermoelasticity benchmarks, compare favorably with a GRU-DeepONet baseline on elastodynamics, and demonstrate apparent invariance to 50% and 200% uniform resampling of the input signals. The architectural idea is coherent and the benchmark results are encouraging, but the resolution-independence claims are not supported by the experiments as currently presented: all training inputs are evenly spaced, the resolution study only subsamples/upsamples the same smooth regular signals, and no quantitative test of arbitrary-time or arbitrary-space querying is reported.

Significance. If fully supported, NCDE-DeepONet would address a real limitation of recurrent DeepONets by providing a continuous-time encoder for load histories and a trunk that can be queried outside the training grid. The paper has several strengths: a clean architectural integration of NCDEs into DeepONet, three different transient PDE benchmarks, consistent hyperparameters across problems, a direct GRU-DeepONet baseline, low reported errors on the test sets, and detailed visualizations in the appendix. However, the headline claims of input- and output-resolution independence are load-bearing and currently outrun the evidence. The missing experiments—irregular sampling with jitter, gaps, or noise, and quantitative evaluation at unseen query points—are well within the scope of the manuscript and would determine whether the title and abstract are justified.

major comments (3)
  1. [§4.3 and Abstract] The input-resolution-independence claim is tested only by subsampling and upsampling the same evenly spaced, smooth signals used in training. The abstract and Section 3.1 promise handling of 'sparse and irregular data,' missing observations, and unevenly spaced samples, but the variable-resolution experiment (Figs. 18–19) changes only the density of a uniform grid over the same smooth boundary conditions. Because the NCDE dynamics in Eq. (3) are driven by the derivative of the interpolated control path, genuinely irregular observation times, missing chunks, or sensor noise would produce control paths never seen in training, and the paper provides no evidence that the terminal latent z(T) remains faithful under such perturbations. I request an experiment with randomly jittered observation times, dropout gaps, and noisy samples, reporting relative L2 errors against the reference for each scenario.
  2. [§3.1 and §4] The output-resolution-independence claim is never quantitatively validated. Section 3.1 states that during training and evaluation, spatiotemporal query points are randomly sampled from 'a discrete set defined by the underlying FEM-generated dataset,' and all reported errors in Eq. (9) are computed on that fixed set. No experiment queries the trunk at unseen spatial locations or times and compares the predictions to reference solutions. To support the claim of arbitrary spatiotemporal querying, the authors should evaluate at previously unseen time instants (e.g., midpoints of the training time grid) and held-out spatial locations, and report the corresponding relative L2 errors.
  3. [§4.3, Figs. 18–19] The resolution-independence experiment reports no quantitative accuracy numbers. The conclusion that the field predictions 'remain in agreement across all sampling rates' is based only on side-by-side color plots. The authors should report the relative L2 error (Eq. 9) for the 50%, 100%, and 200% sampling scenarios against the same reference fields, and ideally also the norm of the difference in the branch latent state z(T) across sampling rates, so that the claimed invariance of the representation can be assessed quantitatively rather than visually.
minor comments (5)
  1. [§2.2] The sentence about ResNet resembling a forward Euler step says '(see Equation 5)', but Equation 5 is the DeepONet formula; this cross-reference should point to Eq. (2) or to a newly introduced ResNet update equation.
  2. [Appendix A.2] The first sentence of Appendix A.2 contains an unresolved cross-reference: 'discussed in Section ??' should be replaced with the actual section number.
  3. [Eq. (3) and Eq. (4)] The notation for the NCDE vector field is inconsistent: Eq. (3) writes \hat f(t,z(t),X(t);\theta)\dot X(t), while Eq. (4) omits the t and X(t) arguments and writes \hat f(z_s;\theta)dX_s. The paper should clarify the dimensions of \dot X(t) and whether the product is a matrix–vector or element-wise product.
  4. [Eq. (6)] In the adjoint equation, \dot X(t) is written as if it were a scalar factor, but X(t) is vector-valued; the authors should specify the indexing/convention so that the transpose and product in Eq. (6) are unambiguous.
  5. [§4.3, Fig. 13] The text states that the top edge has a prescribed constant temperature \theta = 10, while Figure 13 labels it '\theta = C'; these notations should be aligned.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the resolution-independence claim is an architectural property tested empirically, not a fitted input renamed as prediction.

full rationale

The paper's central derivation is architectural and data-driven. The NCDE branch (Eq. 3 and the integral form in Eq. 4) is adopted from the external work of Kidger et al. [6], and the spatiotemporal-trunk coupling with the branch output is the authors' architectural contribution. Input-resolution independence is presented as a property of the construction: discrete observations are first interpolated into a continuous control path, and the latent state evolves under that path, so different discretizations of the same signal produce latent trajectories that depend continuously on the interpolated path. This is a stated design property rather than a result derived from a fitted quantity. The variable-sampling demonstration in Section 4.3 (Figs. 18-19) subsamples and upsamples smooth, uniformly sampled training signals, which is a limited stress test, but the reported agreement is an empirical result and not forced by construction: no parameter is fitted to the 50% or 200% test outputs and then relabeled as a prediction. Self-citations in the reference list (e.g., refs. [10], [26], [36]-[38]) are contextual prior work or baseline-related and are not load-bearing justifications of the central claim; no uniqueness theorem or formal guarantee is imported from the authors' own previous papers. The paper also reports quantitative errors against FEM-generated fields and a GRU-DeepONet baseline, providing external content to the claim. The only identifiable weakness is an untested interpolation-adequacy assumption for genuinely irregular sampling with gaps or jitter, which is a correctness or robustness concern, not circularity. No quoted step reduces by definition to its own inputs, so the appropriate finding is no significant circularity.

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

The paper's contribution is architectural and empirical, not a derivation from first principles. Its central claim rests on standard operator approximation and NCDE well-posedness, on the accuracy of FEM ground truth, and on the modeling choice that one branch vector plus a time-augmented trunk can represent the full field. There are no invented physical entities, and the hand-chosen hyperparameters are few but important.

free parameters (4)
  • Branch and trunk hidden size and depth = hidden size 200, layers 6
    Hand-chosen in Appendix A.1 and identical across benchmarks; the representational capacity of both networks depends on these values.
  • ODE solver and tolerances = Tsit5, rtol 1e-4, atol 1e-7
    Hand-chosen adaptive solver settings determine how accurately the NCDE path is integrated and therefore how much of the resolution-independence result depends on solver error.
  • Input path interpolation = cubic Hermite
    Chooses the continuous control signal from discrete samples; all irregular-sampling claims rely on this interpolation being adequate.
  • Random seed = 2024
    All experiments use a single seed (Appendix A.1), so reported error means and the GRU comparison are point estimates without seed variability.
assumptions (4)
  • standard math DeepONet universal approximation theorem: the trunk-branch inner product can approximate continuous nonlinear operators.
    Section 2.3 relies on the operator approximation theorem to justify the architecture's expressive power, citing [33, 58].
  • standard math Existence, uniqueness, and continuous dependence of NCDE solutions for Lipschitz vector fields.
    Section 3.2 states that a Lipschitz vector field guarantees a unique solution and stable gradient propagation, which underpins training via the adjoint method.
  • domain assumption FEM-generated datasets are accurate ground-truth solutions of the PDEs.
    Sections 4.1 to 4.3 treat finite element results as ytrue in the relative L2 error definition of Eq. (9); no mesh convergence study or experimental validation is provided.
  • ad hoc to paper A single terminal latent state z(T) plus a time-augmented trunk can represent the full spatiotemporal solution field.
    The architecture encodes an entire load history into one branch vector and queries time through the trunk; this factorization is a modeling choice not derived from the PDE structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Time Resolution Independent Operator Learning." pith.science (2026). https://pith.science/paper/LCRRSTCH

@misc{pith2026250702524,
  author       = {Pith},
  title        = {Pith review of: Time Resolution Independent Operator Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCRRSTCH}},
  note         = {Machine review of arXiv:2507.02524}
}
read the original abstract

Accurately learning solution operators for time-dependent partial differential equations (PDEs) from sparse and irregular data remains a challenging task. Recurrent DeepONet extensions inherit the discrete-time limitations of sequence-to-sequence (seq2seq) RNN architectures, while neural-ODE surrogates cannot incorporate new inputs after initialization. We introduce NCDE-DeepONet, a continuous-time operator network that embeds a Neural Controlled Differential Equation (NCDE) in the branch and augments the trunk with explicit space-time coordinates. The NCDE encodes an entire load history as the solution of a controlled ODE driven by a spline-interpolated input path, making the representation input-resolution-independent: it encodes different input signal discretizations of the observed samples. The trunk then probes this latent path at arbitrary spatial locations and times, rendering the overall map output-resolution independent: predictions can be queried on meshes and time steps unseen during training without retraining or interpolation. Benchmarks on transient Poisson, elastodynamic, and thermoelastic problems confirm the robustness and accuracy of the framework, achieving almost instant solution prediction. These findings suggest that controlled dynamics provide a principled and efficient foundation for high-fidelity operator learning in transient mechanics.

Figures

Figures reproduced from arXiv: 2507.02524 by the authors.

Figure 1
Figure 1. (a) Neural Controlled Differential Equation (NCDE) processing of irregular time series. Some data process [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. (a) Gated Recurrent Unit (GRU) processing of irregular time series. Some data process is observed at times [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Schematic of the geometry and boundary conditions of the Poisson example. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Error analysis of the NCDE-DeepONet model on the Poisson test dataset. (a) Distribution of relative errors [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Poisson field predictions at t = 2s for the median error test case. Columns show true field, NCDE-DeepONet prediction, and absolute error for the scalar field u(x, t). 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 t −0.2 0.0 0.2 0.4 0.6 0.8 u = u ( t ) a) 0.00 0.25 0.50…
Figure 6
Figure 6. Figure 6: Time evolution of the Poisson field for the median error test case. True (solid lines) and predicted (dashed [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Error analysis of the spatial-only trunk NCDE-DeepONet model on the Poisson test dataset. (a) Distribution [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Schematic of the geometry and boundary conditions of the elastodynamics example. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Error analysis of the NCDE-DeepONet model on the elastodynamics test dataset. (a) Distribution of relative [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Elastodynamics field predictions at t = 10−5 s for the median error test case. Columns show true fields, NCDE-DeepONet predictions, and absolute errors for displacement components ux and uy. GRU variant replaces the Neural CDE branch network with a Gated Recurrent Uni…
Figure 11
Figure 11. Figure 11: Time evolution of elastodynamics fields for the median error test case. a): Input displacement boundary [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Error analysis of the GRU-DeepONet model on the elastodynamics test dataset. (a) Distribution of relative [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Schematic of the geometry and boundary conditions of the thermoelasticity example. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Error analysis of the NCDE-DeepONet model on the thermoelasticity test dataset. (a) Distribution of relative [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: Thermoelasticity field predictions at t = 250s for the median error test case. Columns show true fields, NCDE-DeepONet predictions, and absolute errors for displacement components ux, uy, and temperature θ. Interpolation capability of NCDE-DeepONet A key advantage of …
Figure 16
Figure 16. Figure 16: Thermoelasticity field predictions at t = 500s for the median error test case. The model maintains accuracy throughout the time evolution, with errors concentrated near the flux boundaries. NCDE effectively learns a continuous representation of the input signals, enab…
Figure 17
Figure 17. Figure 17: Time evolution of thermoelasticity fields for the median error test case. Top: Input heat flux boundary [PITH_FULL_IMAGE:figures/full_fig_p020_17.png]
Figure 18
Figure 18. Figure 18: Input heat flux boundary conditions under different sampling scenarios. Left: Original sampling with 99 [PITH_FULL_IMAGE:figures/full_fig_p021_18.png]
Figure 19
Figure 19. Figure 19: Spatial field predictions at t = 250s for different input sampling scenarios. Despite receiving boundary conditions with 50%, 100%, or 200% of the original temporal resolution, the NCDE-DeepONet produces consistent spatial fields for displacement components (ux, uy) a…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 38 canonical work pages

  1. [1]

    Koric, D

    S. Koric, D. W. Abueidda, Deep learning sequence methods in multiphysics modeling of steel solidification, Metals 11 (3) (2021) 494

  2. [2]

    L. Wu, L. Noels, Recurrent neural networks (rnns) with dimensionality reduction and break down in computational mechanics; application to multi-scale localization step, Computer Methods in Applied Mechanics and Engineering 390 (2022) 114476

  3. [3]

    Deep Learning in Deterministic Computational Mechanics

    L. Herrmann, S. Kollmannsberger, Deep learning in deterministic computational mechanics, arXiv preprint arXiv:2309.15421

  4. [4]

    M. Su, Y . Yu, T. Chen, N. Guo, Z. Yang, A thermodynamics-informed neural network for elastoplastic constitutive modeling of granular materials, Computer Methods in Applied Mechanics and Engineering 430 (2024) 117246

  5. [5]

    Bonatti, D

    C. Bonatti, D. Mohr, On the importance of self-consistency in recurrent neural network models representing elasto-plastic solids, Journal of the Mechanics and Physics of Solids 158 (2022) 104697

  6. [6]

    Kidger, J

    P. Kidger, J. Morrill, J. Foster, T. Lyons, Neural controlled differential equations for irregular time series, Advances in Neural Information Processing Systems 33 (2020) 6696–6707

  7. [7]

    A. H. Ribeiro, K. Tiels, L. A. Aguirre, T. Schön, Beyond exploding and vanishing gradients: analysing rnn training using attractors and smoothness, in: International conference on artificial intelligence and statistics, PMLR, 2020, pp. 2370–2380. 23 NCDE-DeepONet A PREPRINT

  8. [8]

    M. Maia, I. B. Rocha, P. Kerfriden, F. van der Meer, Physically recurrent neural networks for path-dependent heterogeneous materials: Embedding constitutive models in a data-driven surrogate, Computer Methods in Applied Mechanics and Engineering 407 (2023) 115934

Show all 63 references
  1. [9]

    Y . He, S. J. Semnani, Machine learning based modeling of path-dependent materials for finite element analysis, Computers and Geotechnics 156 (2023) 105254

  2. [10]

    Pantidis, H

    P. Pantidis, H. Eldababy, D. Abueidda, M. E. Mobasher, I-fenn with temporal convolutional networks: Expediting the load-history analysis of non-local gradient damage propagation, Computer Methods in Applied Mechanics and Engineering 425 (2024) 116940

  3. [11]

    Y . He, S. J. Semnani, Incremental neural controlled differential equations for modeling of path-dependent material behavior, Computer Methods in Applied Mechanics and Engineering 422 (2024) 116789

  4. [12]

    R. T. Chen, Y . Rubanova, J. Bettencourt, D. K. Duvenaud, Neural ordinary differential equations, Advances in neural information processing systems 31

  5. [13]

    Anumasa, P

    S. Anumasa, P. Srijith, Latent time neural ordinary differential equations, in: Proceedings of the AAAI Conference on Artificial Intelligence, V ol. 36, 2022, pp. 6010–6018

  6. [14]

    K. Lee, E. J. Parish, Parameterized neural ordinary differential equations: Applications to computational physics problems, Proceedings of the Royal Society A 477 (2253) (2021) 20210162

  7. [15]

    C. J. Rojas, A. Dengel, M. D. Ribeiro, Reduced-order model for fluid flows via neural ordinary differential equations, arXiv preprint arXiv:2102.02248

  8. [16]

    Dutta, P

    S. Dutta, P. Rivera-Casillas, M. W. Farthing, Neural ordinary differential equations for data-driven reduced order modeling of environmental hydrodynamics, arXiv preprint arXiv:2104.13962

  9. [17]

    G. D. Portwood, P. P. Mitra, M. D. Ribeiro, T. M. Nguyen, B. T. Nadiga, J. A. Saenz, M. Chertkov, A. Garg, A. Anandkumar, A. Dengel, et al., Turbulence forecasting via neural ode, arXiv preprint arXiv:1911.05180

  10. [18]

    Kidger, On neural differential equations, arXiv preprint arXiv:2202.02435

    P. Kidger, On neural differential equations, arXiv preprint arXiv:2202.02435

  11. [19]

    Morrill, C

    J. Morrill, C. Salvi, P. Kidger, J. Foster, Neural rough differential equations for long time series, in: International Conference on Machine Learning, PMLR, 2021, pp. 7829–7838

  12. [20]

    Morrill, P

    J. Morrill, P. Kidger, L. Yang, T. Lyons, Neural controlled differential equations for online prediction tasks, arXiv preprint arXiv:2106.11028

  13. [21]

    Bleistein, A

    L. Bleistein, A. Guilloux, On the generalization and approximation capacities of neural controlled differential equations, arXiv preprint arXiv:2305.16791

  14. [22]

    J. Choi, H. Choi, J. Hwang, N. Park, Graph neural controlled differential equations for traffic forecasting, in: Proceedings of the AAAI conference on artificial intelligence, V ol. 36, 2022, pp. 6367–6374

  15. [23]

    S. Y . Jhin, M. Jo, S. Kook, N. Park, Learnable path in neural controlled differential equations, in: Proceedings of the AAAI Conference on Artificial Intelligence, V ol. 37, 2023, pp. 8014–8022

  16. [24]

    L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning nonlinear operators via deeponet based on the universal approximation theorem of operators, Nature Machine Intelligence 3 (3) (2021) 218–229

  17. [25]

    L. Lu, X. Meng, Z. Mao, G. E. Karniadakis, DeepXDE: A deep learning library for solving differential equations, SIAM review 63 (1) (2021) 208–228

  18. [26]

    D. W. Abueidda, P. Pantidis, M. E. Mobasher, Deepokan: Deep operator network based on kolmogorov arnold networks for mechanics problems, Computer Methods in Applied Mechanics and Engineering 436 (2025) 117699

  19. [27]

    L. Lu, R. Pestourie, S. G. Johnson, G. Romano, Multifidelity deep neural operators for efficient learning of partial differential equations with application to fast inverse design of nanoscale heat transport, Physical Review Research 4 (2) (2022) 023210

  20. [28]

    Goswami, M

    S. Goswami, M. Yin, Y . Yu, G. E. Karniadakis, A physics-informed variational deeponet for predicting crack path in quasi-brittle materials, Computer Methods in Applied Mechanics and Engineering 391 (2022) 114587

  21. [29]

    S. Garg, S. Chakraborty, Vb-deeponet: A bayesian operator learning framework for uncertainty quantification, Engineering Applications of Artificial Intelligence 118 (2023) 105685

  22. [30]

    Kobayashi, J

    K. Kobayashi, J. Daniell, S. B. Alam, Improved generalization with deep neural operators for engineering systems: Path towards digital twin, Engineering Applications of Artificial Intelligence 131 (2024) 107844

  23. [31]

    S. Cai, Z. Wang, L. Lu, T. A. Zaki, G. E. Karniadakis, Deepm&mnet: Inferring the electroconvection multiphysics fields based on operator approximation by neural networks, Journal of Computational Physics 436 (2021) 110296. 24 NCDE-DeepONet A PREPRINT

  24. [32]

    Goswami, A

    S. Goswami, A. Bora, Y . Yu, G. E. Karniadakis, Physics-informed deep neural operator networks, in: Machine learning in modeling and simulation: methods and applications, Springer, 2023, pp. 219–254

  25. [33]

    L. Lu, X. Meng, S. Cai, Z. Mao, S. Goswami, Z. Zhang, G. E. Karniadakis, A comprehensive and fair comparison of two neural operators (with practical extensions) based on fair data, Computer Methods in Applied Mechanics and Engineering 393 (2022) 114778

  26. [34]

    Ingebrand, A

    T. Ingebrand, A. J. Thorpe, S. Goswami, K. Kumar, U. Topcu, Basis-to-basis operator learning using function encoders, Computer Methods in Applied Mechanics and Engineering 435 (2025) 117646

  27. [35]

    Kumar, X

    A. Kumar, X. Zhi, Z. Ahmad, M. Yin, A. Manbachi, Convolutional deep operator networks for learning nonlinear focused ultrasound wave propagation in heterogeneous spinal cord anatomy, arXiv preprint arXiv:2412.16118

  28. [36]

    J. He, S. Koric, S. Kushwaha, J. Park, D. Abueidda, I. Jasiuk, Novel deeponet architecture to predict stresses in elastoplastic structures with variable complex geometries and loads, Computer Methods in Applied Mechanics and Engineering 415 (2023) 116277

  29. [37]

    J. He, S. Kushwaha, J. Park, S. Koric, D. Abueidda, I. Jasiuk, Sequential deep operator networks (s-deeponet) for predicting full-field solutions under time-dependent loads, Engineering Applications of Artificial Intelligence 127 (2024) 107258

  30. [38]

    J. He, S. Kushwaha, J. Park, S. Koric, D. Abueidda, I. Jasiuk, Predictions of transient vector solution fields with sequential deep operator network, Acta Mechanica (2024) 1–16

  31. [39]

    B. Shih, A. Peyvan, Z. Zhang, G. E. Karniadakis, Transformers as neural operators for solutions of differential equations with finite regularity, Computer Methods in Applied Mechanics and Engineering 434 (2025) 117560

  32. [40]

    C. Yun, S. Bhojanapalli, A. S. Rawat, S. J. Reddi, S. Kumar, Are transformers universal approximators of sequence-to-sequence functions?, arXiv preprint arXiv:1912.10077

  33. [41]

    Zappala, M

    E. Zappala, M. Bagherian, Universal approximation of operators with transformers and neural integral operators, arXiv preprint arXiv:2409.00841

  34. [42]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, A. Anandkumar, Neural operator: Learning maps between function spaces with applications to pdes, Journal of Machine Learning Research 24 (89) (2023) 1–97

  35. [43]

    B. Wang, L. Liu, W. Cai, Multi-scale deeponet (mscale-deeponet) for mitigating spectral bias in learning high frequency operators of oscillatory functions, arXiv preprint arXiv:2504.10932

  36. [44]

    Tancik, P

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, R. Ng, Fourier features let networks learn high frequency functions in low dimensional domains, Advances in neural information processing systems 33 (2020) 7537–7547

  37. [45]

    X. Liu, B. Xu, S. Cao, L. Zhang, Mitigating spectral bias for the multiscale operator learning, Journal of Computational Physics 506 (2024) 112944

  38. [46]

    S. Wang, H. Wang, P. Perdikaris, On the eigenvector bias of fourier feature networks: From regression to solving multi-scale pdes with physics-informed neural networks, Computer Methods in Applied Mechanics and Engineering 384 (2021) 113938

  39. [47]

    Nayak, S

    D. Nayak, S. Goswami, Ti-deeponet: Learnable time integration for stable long-term extrapolation, arXiv preprint arXiv:2505.17341

  40. [48]

    Z. Wu, S. Wang, S. Zhang, S. He, M. Zhu, A. Jiao, L. Lu, D. van Dijk, Tante: Time-adaptive operator learning via neural taylor expansion, arXiv preprint arXiv:2502.08574

  41. [49]

    Tretiakov, X

    S. Tretiakov, X. Li, K. Kumar, Setonet: A deep set-based operator network for solving pdes with permutation invariant variable input sampling, arXiv preprint arXiv:2505.04738

  42. [50]

    Prasthofer, T

    M. Prasthofer, T. De Ryck, S. Mishra, Variable-input deep operator networks, arXiv preprint arXiv:2205.11404

  43. [51]

    S. W. Cho, J. Y . Lee, H. J. Hwang, Learning time-dependent pde via graph neural networks and deep operator network for robust accuracy on irregular grids, arXiv preprint arXiv:2402.08187

  44. [52]

    Bahmani, S

    B. Bahmani, S. Goswami, I. G. Kevrekidis, M. D. Shields, A resolution independent neural operator, Computer Methods in Applied Mechanics and Engineering 444 (2025) 118113

  45. [53]

    Graves, A

    A. Graves, A. Graves, Long short-term memory, Supervised sequence labelling with recurrent neural networks (2012) 37–45

  46. [54]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural computation 9 (8) (1997) 1735–1780. 25 NCDE-DeepONet A PREPRINT

  47. [55]

    R. Dey, F. M. Salem, Gate-variants of gated recurrent unit (gru) neural networks, in: 2017 IEEE 60th international midwest symposium on circuits and systems (MWSCAS), IEEE, 2017, pp. 1597–1600

  48. [56]

    Y . Oh, S. Kam, J. Lee, D.-Y . Lim, S. Kim, A. Bui, Comprehensive review of neural differential equations for time series analysis, arXiv preprint arXiv:2502.09885

  49. [57]

    Dupont, A

    E. Dupont, A. Doucet, Y . W. Teh, Augmented neural odes, Advances in neural information processing systems 32

  50. [58]

    Kobayashi, S

    K. Kobayashi, S. B. Alam, Deep neural operator-driven real-time inference to enable digital twin solutions for nuclear energy systems, Scientific reports 14 (1) (2024) 2101

  51. [59]

    Sharma, V

    R. Sharma, V . Shankar, Ensemble and mixture-of-experts deeponets for operator learning, arXiv preprint arXiv:2405.11907

  52. [60]

    hey, that’s not an ode

    P. Kidger, R. T. Chen, T. J. Lyons, " hey, that’s not an ode": Faster ode adjoints via seminorms., in: ICML, 2021, pp. 5443–5452

  53. [61]

    Matsubara, Y

    T. Matsubara, Y . Miyatake, T. Yaguchi, Symplectic adjoint method for exact gradient of neural ode with minimal memory, Advances in Neural Information Processing Systems 34 (2021) 20772–20784

  54. [62]

    Zhuang, N

    J. Zhuang, N. Dvornek, X. Li, S. Tatikonda, X. Papademetris, J. Duncan, Adaptive checkpoint adjoint method for gradient estimation in neural ode, in: International Conference on Machine Learning, PMLR, 2020, pp. 11639–11649

  55. [63]

    J. L. Ba, J. R. Kiros, G. E. Hinton, Layer normalization, arXiv preprint arXiv:1607.06450. Appendix: Supporting Information A.1. Hyperparameters and network sizes The NCDE-DeepONet architecture employs consistent hyperparameters across all three benchmark problems (transient P...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.