acceptodds
Under review as a conference paper at ICLR 2027

NeuralPDE: Learning Latent PDE Dynamics for Time-Series Forecasting

Abstract

Time-series analysis extracts information from historical observations and predicts their future behavior. Many existing models make predictions only by learning historical data patterns, and the underlying evolution laws, which are often described by partial differential equations (PDEs) or other dynamical rules in real world physical systems, may be overlooked. We therefore propose NeuralPDE, which encodes historical observations to latent functions of an auxiliary variable and learns their evolution through an explicit advection–diffusion–reaction PDE on an auxiliary domain, with a learned readout then decodes the evolving functions to future predictions. NeuralPDE learns PDE-based evolution laws of complex systems in latent space to improve the stability and accuracy of multi-step forecasting. The roles of the transport and diffusion terms in NeuralPDE are theoretically verified through establishing energy dissipation for the linear component. We apply NeuralPDE on real-world time series and NeuralPDE demonstrates competitiveness in time-series multi-step forecasting tasks in comparison with existing models.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, dynamical systems, latent dynamics, neural partial differential equations, continuous-time models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SZiwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces NeuralPDE, a latent dynamics model where the hidden state is represented as a continuous latent function instead of the standard latent vector. The latent function evolves according to an advection-diffusion-reaction PDE, whose coefficients are learned jointly with an encoder-decoder architecture. The PDE is implemented through a Fourier spectral discretization and trained end-to-end. The paper includes theoretical properties of the proposed dynamics and experiments on multiple forecasting benchmarks. The main claim is that latent PDE evolution provides a richer inductive bias than conventional latent ODE models and leads to better forecasting accuracy, particularly for long-term prediction tasks.

Critical strengths

++ The general idea of evolving a latent function rather than a latent vector is intuitively appealing. The ADR decomposition provides a structured inductive bias where advection, diffusion, and reaction have interpretable roles. The technical analysis of the ADR is sound and clearly presented.

++ Strong reproducibility effort: all the hyperparameters, and experimental settings are well documented. The exeperimental protocol is sound and thorough. The code is released anonymously (although I didn’t check the code).

Critical weaknesses

My main concern is about the novelty of the paper. The related work section and positioning is clearly insufficient. There is a broad recent literature on learning PDEs in latent spaces that is not covered:

  • Latent Evolution of PDEs (Wu et al., NeurIPS 2022)

  • Latent Neural PDE Solver (Li et al, 2025)

  • Latent Neural Operator (LNO) (NeurIPS 2024)

  • AROMA (NeurIPS 2024)

After Fourier truncation, the proposed ADR evolution becomes a finite-dimensional ODE solved by a NeuralODE, making the novelty with existing work not clear. It seems to be just another particular parametrization, although elegant, of latent continuous dynamics.

Lack of recent baselines in the experiments: the strongest baseline is a matched Neural ODE. The papers lacks comparison with recent neural solvers (LE-PDE, etc) that are missing in the related work, or at least the authors should justify why these recent baselines are not taken into account.

As acknowledged by the authors, the experimental results are mixed across datasets and configurations. While honest, the paper does not bring insight to the contribution of the proposed approach to forecasting performances, and in what cases it brings a real benefit. The theoretical analysis of the ADR operator, while interessant, is disconnected from the analysis of the forecasting performances.

Questions for authors
  • Could you provide an ablation over the number of Fourier modes?
1 comment

Dear Reviewer SZiw,

As the rebuttal discussion period nears its end, we would like to gently follow up on our responses and the additional experiments presented in the revised manuscript. We remain happy to address any remaining concerns or provide further clarification. If you feel that our responses and additional evidence have satisfactorily addressed your concerns, we would be grateful if you would consider revisiting your rating. Thank you again for your time, thoughtful feedback, and consideration.

Reviewer_Uv89OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes NeuralPDE, which encodes a time-series history as a latent function and evolves it with a learned advection-diffusion-reaction PDE before decoding forecasts. The authors analyze the stability of the linear dynamics and evaluate it on several physical forecasting tasks. NeuralPDE beats a matched latent ODE on some tasks, especially long horizons.

Critical strengths
  1. Representing the latent state as a function over an auxiliary coordinate, and evolving it with a structured PDE, is a novel extension of latent ODEs. This gives the model a richer latent state without adding dynamics parameters as the latent resolution grows, and the transport and diffusion terms provide a physically motivated inductive bias for multi-step forecasting.
  2. The experiments are carefully controlled, with a matched latent ODE baseline, multiple seeds, and leakage controls on the data splits. I also appreciate that the paper reports mixed and negative results openly, which makes its claims easier to trust.
  3. The theory is modest and clearly scoped, and it avoids overclaiming stability for the full nonlinear model.
Critical weaknesses
  1. The empirical case for the added complexity is not very convincing. The gains are somewhat inconsistent across tasks, which is expected in time-series forecasting, but that is why most forecasting papers rely on a large set of benchmarks to show an overall improvement. So some results on standard long-term forecasting benchmarks would be nice to have.
  2. The baselines seem weak: the latent ODE is matched in parameters but not compute, and the other forecasters are less tuned than NeuralPDE. A comparison with PDETime is missing.
  3. The related work section misses closely related forecasting methods that also model the underlying dynamics. While experimental comparisons are not completely necessary, but at least discussing these works and how the proposed method differs from them would enrich that section:
    • Koopa (NeurIPS 2023) also learns time-series dynamics, using Koopman predictors on Fourier-decomposed series.
    • LETS Forecast/DeepEDM (ICML 2025) learns latent dynamics from delay embeddings, and its grounding in Takens' theorem addresses the paper's question of when a latent state can distinguish histories.
Questions for authors
  1. I would be interested to see a compute-matched comparison across all tasks, for example against a wider latent ODE or an augmented ODE with similar runtime, rather than in a single setting.
  2. Adding some results on standard long-term forecasting benchmarks (e.g. https://github.com/thuml/Time-Series-Library) would greatly elevate the paper.
  3. I would be interested in any insight into when the PDE structure helps, for example why NeuralPDE beats the matched ODE on long-horizon cylinder forecasts but not on short horizons or at some sensor counts.
3 comments

Dear Reviewer Uv89,

As the rebuttal discussion period is approaching its conclusion, we wanted to kindly follow up on our responses and the additional experiments provided in the revised manuscript. We would be happy to address any remaining concerns or provide further clarifications if needed. If our responses and additional evidence have satisfactorily addressed your concerns, we would sincerely appreciate your reconsideration of the rating. Thank you again for your time, thoughtful feedback, and consideration.

Thank you authors for your detailed response. Majority of my initial concerns have been addressed and I am upgrading the soundness metric of the review, but I still maintain my original rating of weak acceptance. However I am not opposed to the paper being accepted.

Dear Reviewer Uv89,

Thank you for taking the time to consider our response and additional results. We appreciate your constructive feedback and continued support for the paper. Your comments have helped us improve the paper.

Reviewer_7AVqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that many existing forecasting models rely solely on learning patterns from historical data, while overlooking the underlying dynamical laws governing real-world physical systems. The authors argue that forecasting models should preserve sufficiently rich historical information while learning how that information evolves over time. To address this limitation, it proposes a time-series forecasting architecture that encodes historical observations into latent functions, evolves these functions through an explicit, learnable advection–diffusion–reaction PDE, and decodes the evolved states into future predictions.

Critical strengths
  1. Clear inductive biases and a mathematically motivated architecture. Rather than simply stacking neural-network layers, the model decomposes information propagation, smoothing, and nonlinear transformation into three explicit operators. Compared with unconstrained MLP-based dynamics, this structure may improve data efficiency and forecasting robustness.
  2. Relatively rigorous theoretical analysis with a clearly delineated scope. Beyond establishing energy dissipation for the linear component, the paper analyzes finite-mode representations, spectral formulations, perturbation sensitivity, numerical integration stability conditions, and aliasing.
  3. Thorough reporting of experimental details. The paper provides extensive appendices documenting model parameters, training seeds, data splits, control experiments, and mathematical proofs.
Critical weaknesses
  1. A mismatch between the motivation of preserving historical information and the actual architectural design. The paper motivates its approach by arguing that compressing history into a compact vector may conflate histories that require different forecasts, and therefore proposes a richer function-based representation. In the actual implementation, however, the historical input is first compressed into a 64-dimensional GRU hidden state and then projected into 136 coefficients representing the latent function. Consequently, the contribution appears to concern how to better evolve an already encoded history, rather than how to better preserve historical information.
  2. The advantages of the PDE representation over a conventional ODE have not been sufficiently isolated. The authors seek to demonstrate that latent functions and their PDE-based evolution can model complex systems more effectively than conventional vector-valued ODEs. However, NeuralPDE and the matched ODE use the same number of real-valued state variables in the experiments. Their distinction therefore lies in the parameterization of their vector fields, rather than in the dimensionality of their state spaces. In effect, NeuralPDE is a structured 136-dimensional ODE system, whereas the matched ODE uses hidden layers with a width of only 18. These architectural differences make it difficult to convincingly attribute the reported performance improvements specifically to the physics-inspired advection and diffusion structure, rather than to other factors.
Questions for authors

Please refer to the weakness.

1 comment

Dear Reviewer 7AVq,

As the rebuttal discussion period is approaching its conclusion, we wanted to kindly follow up on our responses and the additional experiments provided in the revised manuscript. We would be happy to address any remaining concerns or provide further clarifications if needed. If our responses and additional evidence have satisfactorily addressed your concerns, we would sincerely appreciate your reconsideration of the rating. Thank you again for your time, thoughtful feedback, and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.