Lie Point Symmetry and Physics Informed Networks

Symmetries have been leveraged to improve the generalization of neural networks through different mechanisms from data augmentation to equivariant architectures. However, despite their potential, their integration into neural solvers for partial differential equations (PDEs) remains largely unexplored. We explore the integration of PDE symmetries, known as Lie point symmetries, in a major family of neural solvers known as physics-informed neural networks (PINNs). We propose a loss function that informs the network about Lie point symmetries in the same way that PINN models try to enforce the underlying PDE through a loss function. Intuitively, our symmetry loss ensures that the infinitesimal generators of the Lie group conserve the PDE solutions. Effectively, this means that once the network learns a solution, it also learns the neighbouring solutions generated by Lie point symmetries. Empirical evaluations indicate that the inductive bias introduced by the Lie point symmetries of the PDEs greatly boosts the sample efficiency of PINNs.

Paper

Similar papers

Peer review

Reviewer KJvK6/10 · confidence 4/52023-07-03

Summary

This study proposes a new loss function for PINNs that imposes the symmetry of PDE.

Strengths

PINN loss comes from the equation, but the proposed loss comes from the property of the equation. The idea is insightful.

Weaknesses

The additional loss increases the computational cost. With the same computational budget, one can increase the number of evaluation points (N_r?) instead of introducing the proposed loss. The comparison might not be fair. The proposed method was only evaluated with the heat equation and Burgers' equation, which are very simple PDEs. Is the symmetry found in more practical and complicated PDEs? Then, is the proposed method useful? The symbols are not unified. In (7), N_l or N_r? Just after (13), N_l=300 is used for data-fit, but the number of points for data-fit is N_0 in (8). In Table 1, "1000" might be wrong.

Questions

With the same computational cost, is the proposed method superior to the vanilla PINN? Is the symmetry found in more practical and complicated PDEs? Then, is the proposed method useful?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The proposed method is limited to PDEs with known symmetry. The generality is unknown.

Reviewer 5AKh6/10 · confidence 3/52023-07-04

Summary

In this paper, a method for finding solutions to differential equations that represent physical phenomena using neural networks is considered. In particular, a method that takes symmetry into account is proposed. Specifically, the authors consider an infinitesimal generator that represents the symmetry of the equation, and constrain the prolongation of the solutions to be orthogonal to the isosurface of the solution expressed in Jet space. This is expected to improve the accuracy of the solutions.

Strengths

As far as I know, the use of the symmetry of the model as an infinitesimal generator, rather than in the form of conservation laws, is certainly new. This enables the use of symmetry for equations that are not derived from the variational principle. The proposed method would be reliable in the sense that the method has a theoretical basis. In addition, the paper is clearly written.

Weaknesses

The computation and implementation of symmetries and the constraints on the jet space based on them are considered to be quite difficult. On the other hand, the improvement confirmed by the numerical experimentation is not significant, so the effect is limited compared to the difficulty of implementation.

Questions

(1) Can the computation of symmetry be automated by using computational algebraic software (e.g., Mathematica, Maple, Sigular)? (2) Although the improvement in numerical experiments is not significant for the cases considered in this paper, can it be effective for systems with many symmetries and conservation laws, such as integrable systems? (3) Perhaps I am misunderstanding something, but is it not sufficient to simply add some additional constraints? For example, does introducing the higher-order derivative of the loss function as additional cost functions (i.e. Sobolev learning) have a similar effect as the proposed method? I suppose that this can be used as an additional constraint because If the loss function is zero for all x and t, then the derivative of the loss function with respect to x and t should also be zero. In this way, it seems to me, it would be easy to create additional constraints without the laborious symmetry computations.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

No potential negative societal impact is expected.

Reviewer dief3/10 · confidence 4/52023-07-06

Summary

The work proposes a generic method to incorporate Lie point symmetry into physics-informed neural networks (PINNs) by augmenting loss function. The method leverages automatic differentiation as other PINNs do because the condition for symmetries is written using differentials. The authors demonstrated the model with symmetry gained better performance than one without symmetry.

Strengths

* The work is based on an established mathematical theory found in Olver (1986). * Once the Lie point symmetries are obtained, it seems quite simple to incorporate them, thanks to automatic differentiation and the PINNs framework.

Weaknesses

* The experiments look insufficient. They compared with and without symmetry but no other state-of-the-art PINN-based models. It makes the contribution of the work look limited because most of the mathematical contributions directly come from Olver (1986), and the results from the symmetry model are not good enough (in particular, Figure 2). They claimed that "our goal is to showcase the effectiveness of using symmetries" (lines 230-231). However, the positive effect of the symmetries is well-known to the community (e.g., Wang et al. 2021a). * The superiority of incorporating symmetries through loss function is unclear. I suggest the authors include data augmentation methods (e.g., Brandstetter et al. 2022a) and equivariant models (e.g., Wang et al. 2021a). * The authors claimed, "Our work presents the foundations for leveraging Lie point symmetry in a large family of Neural PDE solvers that do not require access to accurate simulations" (lines 316-318). However, another part says, "The data for Burgers' equation is obtained using the Fourier Spectral method" (line 299). It makes the reviewer confused about whether the method uses simulation data as supervisory signals. * The contribution of the paper is not clear enough. Is it a new theorem proved, an undiscovered observation found, a state-of-the-art performance, or a novel problem setting? The reviewer recommends the authors clearly state the contribution of the work and show the supporting facts. * The mathematical presentation lacks correctness and clarity. The reviewer found no clear definition of "Lie point symmetry." The definition of $u^{(n)}$ lacks $u$ (see Olver (1986) p.97). It may be by mistake; however, this is an essential part of the theory. Thus the reviewer recommends the authors carefully re-check the manuscript for more mathematical correctness. For instance, if there is no $u$ in $u^{(n)}$, one cannot express the advection term of the Burgers' equation in the jet space. Minor points: * Please add an explanation about $N_s$. * The paper uses "viscosity" for the heat equation, which seems uncommon. It could be called the diffusion coefficient. * In the first equation of Equation (1), $t$ should be defined in an open set (see, e.g., Jürgen Jost "Partial Differential Equations ThirdEdition" (2012) Chapter 1). That's why we need the initial condition.

Questions

* How long do the training and prediction take, respectively? * What is the meaning of "orthogonality" in the paper, e.g. "our symmetry loss encourages the orthogonality of $\mathrm{pr}^{(n)}v$ and the gradient of $\Delta$" (line 237)? The condition contains PDE itself, so the reviewer guesses the use of the terminology could require some explanation.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

1 poor

Contribution

2 fair

Limitations

The limitations are stated sufficiently in the paper.

Reviewer qo6e4/10 · confidence 2/52023-07-13

Summary

This paper proposes to enhance Physics-Informed Neural Networks (PINNs) by incorporating local Lie-point symmetry into them. It is achieved by introducing an additional symmetry loss term, which requires analytic computation of the PDE’s symmetries. This loss term is designed to encourage orthogonality between the PDE equation and the different symmetry transformations. Evaluation is performed on 1D heat and Burgers equations.They show that the incorporation of symmetry gives better performance, especially when the number of sampled points is low. Being unfamiliar with Lie theory, I found the theoretical to be quite challenging. Overall, I found the paper interesting, but a bit short on the experimental side.

Strengths

- The proposed approach is theoretically motivated, and leads to good performance compared to vanilla PINN - It doesn’t need many modifications to the vanilla PINN to implement

Weaknesses

- The proposed method needs to compute the symmetries analytically first, before integrating them into the loss function. Would there be a way to automate this part? - In the considered examples, the better performance obtained with the additional loss terms could have also been achieved by denser sampling, as has been noted. What

Questions

- How does the number of symmetry constraints scale wrt to the dimensionality of the considered PDE? - Could this symmetry loss term be incorporated to other PINN variants too?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

- The experiments are only showcased on 1D PDEs

Reviewer vAL86/10 · confidence 3/52023-07-25

Summary

The paper proposes adding a custom loss function to training physics-informed neural networks (PINNs) to force symmetry requirements when learning solutions to PDEs. A given partial differential equation (PDE) is associated with a Lie group that acts on the space of solutions of the PDE that leaves this space invariant. The custom loss penalizes non-orthogonality between the gradient from the PDE and the symmetry constraint. This forces the updates to the solution to be on the level-set of group invariant solution. The loss is computed using the action of the Lie algebra generators on the jet space of the solutions. The authors show that the additional loss leads to better data efficiency and accuracy in two test case of Heat Equation and Burgers’ Equation.

Strengths

- The idea of adding symmetry constraints to neural networks PDE solvers is a very interesting contribution to the field. - The execution is novel and makes clever use of Lie algebra theory on PDEs. - The exact form of the loss function is clearly motivated. - The experimental results are compelling and convincing.

Weaknesses

- The paper would benefit significantly from a more thorough ablation study. Specifically, an exploration of the effects of incrementally adjusting the relative weighting of the different loss terms may offer crucial insights. It would be nice to analyse the data efficieny in a power low for example, as a function of the loss weights. - The discussion surrounding the two types of loss used in this study (cosine and orthogonality) is not fully clear. While the authors do mention these losses, there is insufficient detail regarding their selection,. - The mathematical notations of the paper are not clear. Overall, I've found the paper quite hard to parse. For example, the notations $\text{pr}^{(n)}[\Delta]$ is not really explained and I could only guess what it means. The Lie algebras are introduced suddently while not referred before. Please re write the introduction of projector using the Lie algebras. - There is an absence of any comparison with baseline models using data augmentation techniques or fully equivariant architecture in term of numerical results.

Questions

- See my point on the power law before. I would plot different power laws based on different triplets of loss weights and show the coefficients. - Please compare to other baselines (including neural network operators and other PINNs) regarding numerical values. There are better ways to compare methods than eyeballing the graphs in Fig. 2 and Fig. 3. In particular because the errors might be very frame dependent. I would give a numerical error by averaging several runs via several seeds over several frames. In the current form, I am unable to fully appreciate the performance to the method. If this point is addressed correctly, I am open to increasing my mark.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Limitations have been well discussed.

Reviewer vAL82023-08-14

Thank you for your answers. ### On the ablation When a new method includes free parameters, it is essential to discuss how they have been chosen and what values were settled upon, even if the impact is found to be minimal. I was unable to locate the final values of each coefficient clearly. You refer to Appendix C (line 275) in the text, but this does not cover that specific detail. Since this new loss is the main contribution of the paper, I am requesting greater clarity on this matter. ### On tables I acknowledge Tables 1 and 2, but I find that there is no clear description of what they exactly represent. I have not seen any explanation of how the test set was selected, so I cannot precisely determine what these numbers mean. I encourage the authors to provide a more rigorous description of their experiments and datasets in the Appendix. ### On the presentation I recognize the challenge of presenting a new (for most of the machine learning community) mathematical concept within the constraints of a short paper. However, complex material can be well presented by being succinct and consistent with notations, something that is not always evident in your paper. For example, you begin by introducing the parametrization of Lie groups by one-parameter subgroups, then proceed to discuss the infinitesimal generators of these subgroups. Later in the results, you start using the term "Lie algebras" without having previously mentioned it. It would be more coherent to introduce cleanly what a Lie algebra is, explain the exponential map, and maintain consistency thereafter. In another instance, you first introduce the prolongation of the action of the Lie group on the jet space and then use the same notation to denote the prolongation of the action of the Lie algebra. This could lead to confusion, and I suggest making the notation and explanation more transparent. Overall I will keep my score and encourage the authors to improve the consistency and clarity of their exposition.

Authorsrebuttal2023-08-20

We thank the reviewer for their answers and especially for pointing out how the presentation of the paper can be improved! **On the ablation:** We performed a hyperparameter search for the coefficients of the loss terms for both models. We found that, similar to PINNs, the model is sensitive to the weight given to the supervised loss for the initial conditions vs the symmetry/PINN loss. The specific coefficient values for the model trained with symmetry loss are: $\beta$=20, $\gamma$=100 for Nr= 500 $\beta$=20, $\gamma$=80 for Nr= 2000 $\beta$=20, $\gamma$=40 for Nr= 10000 The specific coefficient values for the model trained without symmetry loss are: $\alpha$=150, $\beta$=20 for Nr= 500 $\alpha$=150, $\beta$=20 for Nr= 2000 $\alpha$=130, $\beta$=20 for Nr= 10000 We will make sure to include these in the appendix. **On tables:** We have referred to the tables in the Results section for both of the experiments and described their significance, mentioning how they provide evidence that symmetry loss is most useful in low-data regimes. We failed to mention that the test set of 300 initial conditions was selected randomly from the full dataset, and we will mention this in the revision. **On the presentation:** We appreciate the reviewer’s constructive criticisms, especially on how to improve the quality of the presentation of the paper. They will serve to improve the overall quality of our paper in the revision. We believe addressing the two points raised is rather straightforward and would not require a major change.

Reviewer dief2023-08-16

The reviewer appreciates the responses made by the authors. Now it is clear that the method uses no training data and incorporates symmetries behind PDE, which is new. However, the contribution of the paper does not seem sufficient to be accepted at the conference in terms of methodology and experimental significance. The method is a straightforward implementation of Olver (1986) into PINN, and empirical results showed no significance or only marginal improvement. Therefore, the reviewer keeps the score unchanged.

Authorsrebuttal2023-08-20

We thank the reviewer for their response. We emphasize that while developing the theory of symmetry groups of PDEs indeed exists in the mathematical literature, using them to enforce symmetries in neural PDE solvers is a novel contribution. More explicitly, the contributions are listed below, and we will include this list in the revision: 1. Methodology to calculate the group action on the jet space of the PDE using automatic differentiation. 2. Equation for a symmetry loss that enforces arbitrary Lie-point symmetry of the PDE in PINN models. 3. Demonstrate the effectiveness of symmetry relative to increasing the number of sampling points in PINNs.

Reviewer 5AKh2023-08-17

Thank you for the experimental results on the relationship between the number of symmetries and the performance of the method. Based on these results, it seems that the proposed method is very effective for equations with a large number of symmetries and conservation laws, such as integrable PDEs. In particular, if the method can be combined with computational algebra so that a large number of symmetries can be easily, or automatically, handled, it can be a very powerful method. However, as the number of symmetries in the additional experiment is limited, there remains a concern that as the number of symmetries increases, the performance improvement may saturate. So I will keep my score this time; however, if additional experiments that show that the performance improvement would not saturate are provided, I am happy to increase the score.

Authorsrebuttal2023-08-20

We thank the reviewer for acknowledging the significance of the proposed method. Unfortunately, current PINN models that can take as input different initial conditions fail for higher dimensions. This is due to inherent difficulties in training PINN models, the cost of calculating gradients, and other gradient pathologies that have been studied previously [1]. Given these limitations, it isn’t feasible to showcase the performance of the proposed algorithm for high-dimensional or complex systems. We hope that further developments in PINN model will allow us to exploit their symmetries with the proposed algorithm. [1] Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient pathologies in physics-informed neural networks, 2020b.

Reviewer 5AKh2023-08-21

Thank you so much for your response. Although not completely, to some extent, my concerns have been addressed and I have increased the score.

Reviewer KJvK2023-08-19

Thank you for your response.

Thank you for your response and additional experiments. As I am concerned, simply using much more points leads to a better result. While introducing symmetry is always effective, I would like to see curves of the computational cost vs performance for a fair comparison. In other words, with the same computational budget, is introducing the symmetry loss more effective than increasing the number of points? > Finally, a common question is about the automatic derivation of symmetries from the PDE. This is indeed possible – computational algebraic software can calculate symmetry groups for a given PDE. Symbolic programs in MACSYMA, REDUCE, MAPLE, and Mathematica have been developed to find the equations for the infinitesimal generators. It is not easy to agree on this point without a demonstration. If you could demonstrate your method with such a procedure, I would give it a better score.

Reviewer qo6e2023-08-21

Thank you for addressing my questions. However, due to the limited numerical experiments, I will keep my rating

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC