The Challenges of the Nonlinear Regime for Physics-Informed Neural Networks

The Neural Tangent Kernel (NTK) viewpoint is widely employed to analyze the training dynamics of overparameterized Physics-Informed Neural Networks (PINNs). However, unlike the case of linear Partial Differential Equations (PDEs), we show how the NTK perspective falls short in the nonlinear scenario. Specifically, we establish that the NTK yields a random matrix at initialization that is not constant during training, contrary to conventional belief. Another significant difference from the linear regime is that, even in the idealistic infinite-width limit, the Hessian does not vanish and hence it cannot be disregarded during training. This motivates the adoption of second-order optimization methods. We explore the convergence guarantees of such methods in both linear and nonlinear cases, addressing challenges such as spectral bias and slow convergence. Every theoretical result is supported by numerical examples with both linear and nonlinear PDEs, and we highlight the benefits of second-order methods in benchmark test cases.

Paper

Similar papers

Peer review

Reviewer XYBU6/10 · confidence 3/52024-07-04

Summary

The authors analyze the NTK perspective for PINNs for non-linear PDEs. Previous considerations derived for linear PDEs fall short for non-linear ones and the authors attribute this difference to the non-vanishing Hessian term. Therefore, they suggest to use second order methods and show to be able to converge faster in one experiment. Second order method are shown to be useful also for linear PDEs as they alleviate the spectral bias of the NTK.

Strengths

- the paper is well written and flows very nicely - I found the analysis about the NTK in non-linear regime elegant and the experiments clearly support the findings

Weaknesses

- the different behaviour of the NTK for non-linear PDEs is not very surprising - it is also not very surprising that second order method can work better, but in practice they come with significant shortcomings. Even though the method is shown to be faster in one experiment, no rate is derived so it might not be true in general. However, the applicability of Theorem 4.2 is just sketched (lines 218-222)

Questions

- I would like the authors to elaborate on the applicability of the second order method. In particular, I found the explanation in lines 218-222 a bit hand-wavy. It seems like the hypothesis of Theorem 4.2 (e.g. J(t) being full-rank) are very hard to check in practice - No guidance or intuition is provided for practitioners on when it might be convenient to use the second order method, given that the method is expensive (the authors only mention general well-known considerations for second order methods in the limitations paragraph)

Rating

6

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

Limitations are discussed. However, as mentioned in the question section, I think the authors should discuss in more details the scalability of the proposed second order approach and on when it might be convenient to use it.

Reviewer ocfJ7/10 · confidence 5/52024-07-08

Summary

This paper studies the training dynamics of PINNs, especially for the nonlinear PDEs. The authors find that the previous recognized NTK viewpoint is not applicable to nonlinear PDEs, although it holds for linear PDEs. Therefore, the global convergence of gradient descent on nonlinear PDEs may not be guaranteed. Moreover, the imbalance of singular values of Gram matrices, which also occurs in linear PDEs, results in slow convergence. To address this issue of spectral bias, the paper suggests using second-order methods for parameters update. Experimental results show that LM Newton’s method can achieve lower training loss compared to Adam and L-BFGS.

Strengths

The paper is well-written and provides both theoretical and empirical analyses. The failure of NTK approach on nonlinear PDEs has not been highlighted in previous studies of the global convergence of training PINNs with gradient descent. The second-order method is essential in smooth problems and deep learning training. The authors highlight the effectiveness of Newton’s method in balancing singular values of the training dynamics.

Weaknesses

(1) the failure of NTK is similarly investigated in some previous works (e.g., https://www.sciencedirect.com/science/article/pii/S016727892300341X), where they found that the Gram matrix does not consistently converge in some cases. Therefore, the observation seems to be not novel. They also pointed out that NTK may also work for some nonlinear PDEs. (2) the second-order methods for modifying the singular values of Gram matrix of training dynamics are also not new. (3) the Hessian inverse is quite expensive in practice. (4) regularizing the Hessian may also be intractable in practice, for high dimensional problems, e.g., training PINNs.

Questions

I have the following questions: (1) In the previous work (e.g., https://www.sciencedirect.com/science/article/pii/S016727892300341X), it was found that NTK approach fails in solving some PDEs. Besides linear PDEs, it is still possible that NTK works for some nonlinear PDEs. However, in your paper, you exclude the special situation. It seems to be more restrictive than the published work. (2) Although the second-order methods enjoy very good theoretical properties, deep learning community typically prefers first-order methods due to their computational efficiency. In practice, Newton’s method is more computational expensive even with some inexact technique (e.g., Krylob subspace, conjugate gradient, and LBFGS). However, your theorem of global convergence (Theorem 4.2) does not apply to these inexact methods. Based on your result (which I believe built upon some previous works), can you extend your findings to more practical inexact Newton’s method (e.g., replacing Hessian inverse by Krylov subspace method or Quasi-Newton’s method of BFGS or LBFGS)? (3) In your experimental results, the LM method performs significantly better than LBFGS. This is surprising, as LBFGS or BFGS asymptotically approaches the exact Hessian under certain conditions. Moreover, in case where the totally number of grid points are large and the batch is small in a stochastic setting, inexact and Quasi Newtons method (e.g., Hessian averaging methods including BFGS and LBFGS) should hold the asymptotic convergence. Therefore, intuitively, I would expect LBFGS to perform at least comparably with your LM method. Am I correct? Can you explain why LBFGS performs poorly according to your results, although it can approximate Hessian and extract Hessian information?

Rating

7

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes

Reviewer 4Qrf7/10 · confidence 3/52024-07-09

Summary

The paper studies the NTK of NNs trained on non-linear PDEs, showing that they exhibit different behaviours compared to standard analysis of NTKs. The paper then discusses the issue of spectral bias that arises from first-order methods, showing that they can be alleviated by the use of second-order methods.

Strengths

- The paper presents an interesting analysis of a common tool in NNs and PINNs, and presents an explanation why second-order methods (which are already used in PINNs to some extent) works better than first-order methods. - The paper provides both theoretical and empirical justification for the various claims, and is well-organised in that manner.

Weaknesses

- The sections could be a bit more coherent. For example, the paper brings up the properties of the NTK in the nonlinear PDE case, but then provides less link of these properties of how it affects the convergence in terms of the spectral biases. The LM algorithm is also brought up as a second-order optimisation method, however it may warrant more description as to why it is introduced or how it differs from existing second-order methods such as LBFGS. - Explicit mention of LM algorithm's runtime could be mentioned for completeness.

Questions

- Already in PINNs, there are many works that uses NTKs in loss function scaling [1], collocation point selection [2], analysis of PINN architectures [3], and more. How would the insights in the paper be able to address the points raised in these papers, and how would they affect these proposed methods? - Is Theorem 4.2 general enough to be applied to regular NNs as well? How does the result compare to existing theoretical works on second-order methods in NNs or general optimisation problems? [1] Wang et al. When and why PINNs fail to train: A neural tangent kernel perspective. [2] Lau et al. PINNACLE: PINN Adaptive ColLocation and Experimental points selection. [3] Wang et. al. On the eigenvector bias of Fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

4

Limitations

Limitations suggested are of the LM algorithm which is adequate.

Reviewer vTvV7/10 · confidence 4/52024-07-13

Summary

In this paper, the theory of the Neural Tangent Kernel(NTK) in the case of solving nonlinear partial differential equations using PINNs is investigated in detail. In particular, it is shown that typical results of the NTK framework do not hold when the simple gradient descent method is employed due to the worse behavior of the Hessian matrix compared to the linear cases. In contrast, when second-order methods are employed, it is theoretically proven that the training of the neural networks is efficient.

Strengths

This paper theoretically investigates the behaviors of learning dynamics for PINNs, which are known to be difficult to train. This paper provides a theoretical guarantee of the effectiveness of second-order optimization methods for training PINNs for nonlinear partial differential equations. Im my opinion, this is a significant result, which may lead to applications of PINNs to practical problems that have been inapplicable due to the training difficulties.

Weaknesses

My concern about this paper is in the increase of the computational complexity of second-order methods; however, this concern has already been discussed by the authors in the paper.

Questions

In the numerical experiments, it seems that not so large neural networks are employed. Is it expected that neural networks of this size behave like the results of the theory?

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

There seems to be no problem.

Reviewer ocfJ2024-08-09

Answer to the rebuttal

Thank you for the clarification. My concerns are well addressed. I would like to raise my score. To further improve the quality of the paper and make clear claim, I hope authors can include the above related discussions (e.g., the convergence of gram matrices holds for nonlinear PDEs when s>1/2, and the extension (although may fail in practice) to other quasi-newton methods).

Authorsrebuttal2024-08-13

We are glad that our reply effectively addressed the reviewer's concerns and we deeply appreciate the increase in score. It will be our pleasure to include in our paper the discussion above.

Reviewer vTvV2024-08-13

I appreciate the authors for the detailed reply. Because I have already given a high score, I will keep my score.

Authorsrebuttal2024-08-13

We are pleased that our response met the reviewer's expectations and feedback.

Reviewer XYBU2024-08-14

I would like to thank the authors for their clarifications

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC