acceptodds
Withdrawn from ICLR 2027

WINN: A PDE Residual Loss-Free Network for Homogeneous Long-Time Wave Problems

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Physics-informed neural networks (PINNs) struggle to solve wave equations over long temporal domains. In this work, we investigate the sources of this failure and introduce WINN (Wave-Informed Neural Network), a novel neural architecture that incorporates the analytical structure of the wave equation directly into the model. WINN satisfies the homogeneous wave equation with constant wave speed exactly by construction, eliminating the need for a PDE residual loss and allowing training to focus solely on the initial and boundary conditions. We show through extensive experiments that WINN remains accurate across long-time wave problems with different frequencies, spatial dimensions, wave speeds, and domain geometries, achieving – lower relative error than the best-performing baseline. WINN achieves training times that are slightly shorter than Vanilla PINN in 2D and 1.7 faster than FBPINN in 3D. These results demonstrate that incorporating the governing PDE structure directly into the neural representation can substantially improve the accuracy and robustness of neural solvers for long-time wave dynamics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-informed neural networks, neural PDE solvers, wave equations, scientific machine learning, long-time dynamics, structure-preserving neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_kgpnOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper considers the standard wave equation and propose a PINN structure by composing the triangle functions and an MLP, followed by a integrand, whose parameterization exactly satisfies the wave equation. Experiments are conducted on such homogeneous wave equation with constant speed, and the method is expected to achieve lower MSE than unconstrained method.

Critical strengths
  1. The idea of explicit wave equation parameterization for a PINN seems interesting and clever.
  2. The presentation is generally easy to follow.
Critical weaknesses
  1. The applicability is really narrow. The method can only evaluate homogeneous wave equation with a known solution and parameterization. The applicability to more complex PDEs, which is the true important challenge, is not straightforward and is not presented in the paper. The paper only justifies a simple combination of PINN and PDE solution family given that the exact solution is known. For general high-dimensional and complex PDEs, the method can not be evaluated since the exact solution family is not known. So how can this method be applied in real scenarios?

  2. From this perspective, the experiments are generally flawed. The method WINN has already encode the solution function exactly in its parameterization, so the performance improvement is unfair and not convincing, in the sense that whether the authors provided a new type of PINN structure that is widely applicable and generalizable to other problems.

Overall, while the idea seems interesting, the underlying solution function used in the parameterization makes the method not a generally sound and convincing structure for the PINN setting.

Questions for authors

I have no further questions.

Reviewer_teP6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose a neural network architecture which enforces the wave equation in an homogeneous medium as a hard constraint. Training such a neural network to solve a wave IBVP (in a homogeneous medium) requires only minimizing losses related to the initial condition and to be boundary conditions. This construction is, then, applied to model wave propagation for long temporal horizons (which is a particularly difficult problem for PINNs).

The architecture is tested for 2D/3D, low- and high-frequency scenarios, and for non-rectangular domains. The results are compared to established models (Vanilla PINNs, FB-PINNs, FF-PINNs and PINNs with causal training).

Critical strengths

The paper is well-written, the presentation is clear, and the comparison results are good. The main idea of the paper is interesting.

Critical weaknesses

In my point of view, there is a single critical weakness, but this is an important one: the method is limited to a very narrow set of problems. Indeed, the authors propose a method which deals only with homogeneous velocity fields, which, by the way, is the simplest scenario for wave propagation. I have some concerns about the scientific relevance of such a method.

Questions for authors

Have you tried to adapt your method to deal with non-homogeneous media? (Not necessarily imposing the wave equation as a hard constraint in the way you did, of course, but at least using your construction as some kind of "weak" inductive bias by adding some other features to learn the "non-homogeneous evolution", maybe.) Having some results in non-homogeneous media would dramatically increase the scientific value of the paper.

Reviewer_uCLjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The pitch of the paper is that PINNs struggle with long-time wave problems, and that this difficulty can be avoided by building the wave equation into the network. The authors use a known analytical representation to ensure that WINN satisfies the PDE by construction, so training only needs to fit the initial and boundary conditions.

Critical strengths

The main strength is that WINN satisfies the governing PDE exactly without requiring a PDE residual loss. The reported improvements over PINN baselines are substantial across all four problems, including different frequencies, spatial dimensions and geometries. The experiments also investigate the failure of conventional PINNs as the temporal horizon increases.

Critical weaknesses

The main claim concerns long-time wave problems, but only Figure 1 systematically varies the temporal horizon TT, and this is for one standing-wave example. The other experiments seem to use a single TT per problem. So how do we know that the improvements in Table 2 persist as TT increases for the different problems?

The writing also leaves important parts of the method, particularly the Fourier encoding and explicit training losses, in the appendix, even though Table 3 shows how much accuracy depends on Fourier features. So this is kind of badly written that much of the important details are in the appendix.

In general the experiments are restricted to simple linear PDEs and hence thats not much of an evidence for a new way to do PINNs.

Questions for authors

RC1

Table 2 gives the main results, but does not state the temporal horizon TT. These values are only given in Appendix C.1: T=100T=100 for the three 2D problems and T=50T=50 for 3D. This choice of TT should be stated in Table 2. More importantly, can the authors vary TT on these other problems, as in Figure 1? Every test case needs to have a version of Figure 1 for it. Otherwise it is not clear how robust the long-time claim is.

RC2

All comparisons are with other neural solvers. Could the authors include a conventional numerical solver on the same problems and compare error at matched computational cost? The current results show that WINN is better than the PINN baselines, but not whether it is competitive with established methods for this equation. Given these PDEs are simple, that argument becomes important.

RC3

The main comparisons use five hidden layers of width 64 for WINN and the non-decomposed PINN baselines. Do the differences persist if the networks are made smaller or larger? A controlled width study for WINN and FF-PINN would show whether the gains depend on this particular model capacity.

RC4

Table 3 shows that removing Fourier features increases error from 4.21% to 100.04%, yet the encoding itself is defined only in Appendix B.1. The explicit training losses are also in Appendix C.3. These are central to understanding the method and should be explained in the main text.

RC5

The method is tested only in two and three spatial dimensions. PINNs in low dimensions is just not interesting as the whole point of the PINN technology is to remove the curse-of-dimensionality of PDE solving. Could the authors try a higher-dimensional synthetic wave problem, say 10D or even 50D? A proof of PINN modification being useful has to be in terms of the advantage persisting across scaling of the dimension.

RC6

Wave PDEs are fundamentally simple/linear PDEs. Is there a version of this idea that can be useful for other/non-linear PDEs? Its kind of unclear why effort must be invested into improving performance for simple PDEs where standard non-neural methods are already extremely well-tuned.

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

Related papers

Discussion (0)

Sign in to comment.