acceptodds
Under review as a conference paper at ICLR 2027

Governing Equation Discovery from Data Based on Differential Invariants

Abstract

The explicit governing equation is one of the simplest and most intuitive forms for characterizing physical laws. However, directly discovering partial differential equations (PDEs) from data poses significant challenges, primarily in determining relevant terms from a vast search space. Symmetry, as a crucial prior knowledge in scientific fields, has been widely applied in tasks such as designing equivariant networks and guiding neural PDE solvers. In this paper, we propose a pipeline for governing equation discovery based on differential invariants, which can losslessly reduce the search space of existing equation discovery methods while strictly adhering to symmetry. Specifically, we compute the set of differential invariants corresponding to the infinitesimal generators of the symmetry group and select them as the relevant terms for equation discovery. Taking DI-SINDy (SINDy based on Differential Invariants) as an example, we demonstrate that its success rate and accuracy in PDE discovery surpass those of other symmetry-informed governing equation discovery methods across a series of PDEs. Additional results further indicate that our method exhibits strong robustness to dataset and symmetry noise, significant potential for collaborating with symmetry discovery approaches to solve high-dimensional systems with unknown symmetries, and the ability to serve as a universal plug-and-play module compatible with diverse equation discovery methods.

Primary area:
learning on time series and dynamical systems
Keywords:
equation discovery, symmetry, Lie theory, equivariant neural network

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

The paper solves the problem of discovering symbolic forms of PDEs from observational data. Existing search-based and regression-based methods (like SINDy) require defining a dictionary of terms to be searched over. The authors propose to instead discover symmetries (or use existing knowledge of them) to first construct a basis/dictionary to be used. This is all done on a linear level by finding the symmetry lie algebra, which is then derived into a set of differential invariants. The method is compared against SINDy and EquivSINDy-r and more.

Critical strengths

The theoretical basis of the method is well grounded, with classical results from Olver, and the dictionary of differential invariants can be used in many broad ways within discovery frameworks. The paper is easy to read, providing plenty concrete examples and extensive evaluations, particularly the additional experiments in the appendix.

Critical weaknesses

The proposed method only becomes helpful for problems where the symmetries are explicit, equations admit them, and there are sufficiently many of them. It is unclear how useful these are when applied to more real-world discovery problems.

The scope of the paper is quite limited: the problem of discovery of symmetries is not tackled here, and theoretical results are all classical results in the field of DE symmetries. It is a nice application, highlighting the usefulness for an ML community, and should be seen as a computational study primarily. However, a lot of interesting experiments and ablations appear in appendix only -- the authors should do a more thorough job of communicating the experimental evidence in main body.

Questions for authors
  1. Why is Section 3 even a section? This appears to be at most a paragraph, if that.

  2. Could authors expand on the logic in Appendix I.2 or remove it altogether? When the true PDE order is unknown, I am not convinced that the proposed approaches are generally justified. In particular, I do not believe that the order can be inferred from the highest order in the infinitesimal generators, as different-order PDEs can admit the same generators. E.g. Table 2 itself is a counterexample: KdV, KS and Burgers share exactly the same generators but have orders 3, 4 and 2. Similarly, high-frequency spectral decay does not generally determine the differential order without additional assumptions.

  3. Another thing -- the paper multiple times mentions "losslessly", however taking even the reaction–diffusion example, the 5 invariants in Appendix H.3 look incomplete. With 14 variables and two rotations I’d expect about 12, and the set has no first-derivative terms at all, e.g. ∣∇u∣2+∣∇v∣2|\nabla u |^2 + |\nabla v |^2. Is the set meant to be complete? If not, the “lossless” claim doesn’t hold here.

  4. The reported success rates and RMSE appear to consider only the learned coefficients WW, excluding coefficients fixed within the differential invariants. For example, in Appendix H.2, δ=0.5\delta=0.5 introduces a 50% error in the nonlinear coefficient of the Burgers equation, yet Table 5 reports 70% success and an RMSE of only 1.94×10−31.94\times10^{-3}. This still counts as a success with a tiny RMSE, while EquivSINDy-r’s RMSE does include that coefficient. Figure 5 δ=0.5\delta = 0.5 actually shows DI-SINDy with the largest long-term error. Could you report the error of the full recovered equation, including the LHS coefficients? Otherwise Tables 3 and 5 don’t compare the same thing.

  5. Line 306 -- should be Sivashinsky

  6. Line 54 -- “Each symmetry corresponds to a conserved quantity” is only true for variational symmetries

Reviewer_FGbKOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes the use of differential invariants in partial differential equation (PDE) identification. When a symmetry group of a differential equation is known a priori, either fully or partially, the PDE becomes a relation between these corresponding invariants. Thus, once can form a minimal correct basis on which to perform the identification task. The paper develops DI-SINDy as the corresponding symbolic identification method. The paper begins with introduction and related work sections, reserves Lie algebra discussion for the appendix, and introduces the method in section 4 (computation of the differential invariants and the identification method). Computational experiments are presented or outlined in Section 5, with concluding remarks in section 6.

Critical strengths
  • The method as presented is sound, the development is well-founded in classical PDE invariant theory.
  • The range of examples shows that the method can be used in a broad setting, even at times when the symmetry group is not fully known, and with a variety of compitational methods.
  • Comparisons show that other symmetry-informed methods, which impose symmetry differently, may lag in performance compared to a DI-inspired method in certain settings.
  • DI methods can yield nonlinear features, which can be a significant advantage compared to other forms of basis elements, since they are hard to approximate.
Critical weaknesses
  • Infinite-dimensional symmetries appear in very simple equations, e.g. the heat equation, and it is unclear what happens then; does the method extend, or does it fundamentally fail. Also, does the method extend to symmetries of the first jet space. A discussion regarding the boundaries of this method (a) in terms of PDE theory, and (b) in terms of computational capability, with respect to symmetry group estimation methods, would help qualify the settings in which the method can be reliably used.
Questions for authors
  • can the authors comment on where the method can be extended theoretically, and where theory prohibits this extension? (with respect to infinite-dimensional, or genuine contact actions?)
  • The DI computation hinges on the correct identification of the point symmetry group. The failure modes of the corresponding methods for symmetry discovery should be discussed in the main text (in addition to their performance and deterioration under noisy observation).
Reviewer_1cQeOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method that uses Lie symmetry-based differential invariants to improve symbolic equation discovery for PDEs. In the case where the symmetries are known, it first requires the experimenter to hand-derive the differential invariants of the symmetries, and then use these differential invariants as constraints when searching for the terms for the equation. In the case where the symmetries are not known, it first calls a data-driven symmetry discovery method to obtain the symmetry generators, and then the rest of the pipeline follows.

The primary contribution is that it uses symmetries to aid the process of searching the terms in SINDY for PDE. The authors' experiments show that their proposed method (DI-SINDY) achieves the best performance compared to SINDy, EquivSINDy-r on some classical PDE benchmarks.

Critical strengths
  1. This work is novel as an assembly. To my knowledge, it might be the first work that utilizes Lie symmetry-based differential invariants on symbolic equation discovery method. The method's validity is also apparently shown in its experiments comparing DI-SINDy to two other SINDy methods. The improvement of the method, compared to the baselines, is significant.

  2. The work has a clear and consistent use of relevant mathematical knowledge. It's well-written and understandable. The mathematical concepts from Olver's book are duly included and presented in a logical order.

Critical weaknesses
  1. It's heavily relying on either knowing the exact symmetries of the underlying PDE behind the observed trajectory data or having access to a strong symmetry discovery method. 1) Assuming the knowledge of exact symmetries (their analytic forms) of the underlying data is a strong prior that, in some sense, has solved half of the equation discovery problem already, since the structures on top of the equation manifold are already known from the symmetries. This thus reduces the significance of the paper's contribution. 2) If we don't assume knowledge of the symmetries, then the method heavily depends on the robustness of the symmetry discovery (abbreviated as SD) method. The SD method the authors used (Hu. et al, 2025) is by itself a symbolic method (as shown in the beginning of "4. Method" section of the paper), which provides a natural leverage of obtaining the exact, analytic expressions of the symmetries. Two disadvantages follow: a) if the discovered symmetry generators are in neural form, then the differential invariants found might not guarantee to help discovering the PDE terms, and the main concern is that b) using a symbolic SD method means it might not be able to correctly discover the symmetries when the true symmetries are outside the expressivity of the terms available - in the case where the underlying PDEs and its symmetries are genuinely unknown (i.e. we have no access to its ground-truth symmetries to check, but rather only its trajectory data), there is no robust way to test if the library is expressive enough. Based on the above two points, the reliance on symmetry - especially their symbolic forms - is a major bottleneck of this entire method.

  2. The pipeline depends on two other strong prior knowledge. 1) The work does not show its applicability on PDEs that have no symmetries at all. In another word, "assuming that the underlying trajectory data's governing equation has a symmetry" is a strong assumption that can be rarely made out-in-the-wild. When only a set of trajectory data are given, there is no way to know if symmetry exists. If the SD method is applied on top of a set of trajectory data that have no Lie symmetries at all, the entire pipeline (both DI-SINDy and the SD method) most likely wouldn't be working right. 2) The work assumes knowledge of the differential order of the PDE. In appendix I.2, the authors do acknowledge this weakness and claims that, in practice, a larger n is chosen to ensure all ground-truth terms are included. However, this doesn't seem to be supported by evidences. In the same appendix, the authors suggest that this can be included in future works, but the differential order is a key component in both Lie symmetries and PDEs, which means the postponement of the implementation to future works isn't very convincing.

  3. Potential numerical instability. The central difference method is most likely a noisy method for estimating derivative data, which accumulates numerical errors as the differential order increases.

  4. Expressivity. Despite SINDy is a canonical method for explicit/symbolic discovery of equations, its use of library poses an intrinsic limit on the expressivity, especially in high-dimensions (such as a PDEs with >50 -dimensional state space and >50 number of independent variables) and when the true underlying equations are highly nonlinear. However, this is potentially a limit of SINDy instead of the paper's method itself, yet this would make the paper's "plug-and-play" claim less contributive in the sense that there isn't much to plug into since SINDy isn't really the mainstream way of parametrizing an equation in relevant engineering tasks.

Questions for authors
  1. What happens if the underlying PDE has no symmetries? What would this proposed pipeline give us?
  2. Can you show some experiments regarding the differential order of the PDE?
  3. How does the numerical noise from central difference influence the performance of the pipeline, especially with much higher orders?
  4. How much can this pipeline scale? Can you do some high-dimensional examples such as PDEs with >50 number of independent & dependent variables?
  5. What if the symmetry discovery method discovers symmetry generators in neural form? What if the method proposed in Hu et al, 2025 does not have a library that is expressive enough to capture the symmetries? How would DI-SINDy work with a different symmetry discovery method?
5 comments
Reviewer_1cQeOpenReview ↗

Response to authors

Overall, although I appreciate the experiments the authors conducted, I do not think the responses fully address some of my main concerns despite some concerns are resolved.

  1. In your response to W1, I agree with that a usable subgroup or just a few symmetry directions can still supply a valid invariant representation and provide useful priors to constrain the search space of equation discovery - however, this is a claim on completeness of the discovered symmetry group instead of the quality of the discovered symmetry group, and my original concern is not on the completeness, but on the quality, i.e. if the upstream symmetry discovery method produces a very bad set of symmetry generators, then essentially the authors' proposed method would fall apart because no useful symmetries are available.

I also agree with that observations are still needed when identifying the generating equation even if the symmetry prior constrains an equation family, but that doesn't counter my point: "having the right, good-quality symmetry of the underlying equation" is a strong prior that makes the method fall into a circular logic, in which we begin with knowing the symmetry (i.e. we know a faithful representation of the structure solution space of the equation we're discovering), and then use the symmetry to discover the equations to show that its performance is improved. The part about it that is circular is that knowing symmetry is a strong prior that has basically solved a big chunk of the equation discovery task you're doing, and, in the real & raw data regime (out in the wild), one cannot expect to know the symmetries of the underlying equation already (only in some physically interpretable sense, such as rotation, translation, scaling symmetries etc. - but they are only a very small subset of symmetries). The authors' way to tackle this is to use symmetry discovery method, but then how do you know the symmetry discovery method is gonna give you good symmetries? To be more concrete: LieNLSD uses a symbolic library - but what if LieNLSD's library is not expressive enough to produce faithful symmetries from data that you can use in the method of equation discovery the authors proposed?

The experiment of removing phase rotation thus does not resolve my concern because it only supports that "not knowing the complete symmetry group is okay" but does not support that "using low-quality symmetries can still help with discovery".

  1. In response to the authors' response to W2(1)/Q1, if no useful symmetries are returned and that the proposed method degrades into an ordinary equation discovery, doesn't this serve as evidence to my W1 claim that the method heavily depends on the upstream symmetry discovery method returns a good (if any) set of symmetries? The authors' proposed method is on using symmetries on equation discovery, not on symmetry discovery itself. And, if on some data the SD method returns no symmetries and it's just vanilla equation discovery, then does that mean the proposed method fail in the case when the underlying data has no symmetries?

  2. I appreciate the authors for conducting the differential order experiment in response to W2(2) and Q2. This fully resolves my concern regarding the differential order only. I also appreciate the experiment of different noise levels - it fully resolves my concern regarding differential noise only. I appreciate the authors for pointing out the E2E and PySR tasks, yet these are still symbolic equation discovery methods from benchmark PDE data and it doesn't establish compatibility with common neural pDE modeling approaches such as Fourier neural operators, neural surrogate models, or physics-informed neural networks, nor in a complex, measured engineering system. I think the scope of the plug-and-play claim is still only available in the explicit symbolic equation-discovery method regime but not in universal engineering applications and learning architectures - I do agree with that this is a limit of symbolic method itself instead of the paper's proposed method, but it still worths mention because the scope it's operating in determines its actual engineering impact.

  3. I appreciate the high-dim PDE experiment. It resolves the concern on high-dimensionality only.

  4. The concern on neural symmetry generators is not on its mathematical compatibility - it's on the numerical stability and implementation feasibility. I do not think the authors claim of the mathematical compatibility of the proposed method with neural symmetry generators is enough to resolve my concern because numerical issues might arise in implementation.

  5. I appreciate for pointing out the LieGAN generators, which indeed resolves my question of how it would work with a different symmetry discovery method. However, it still doesn't resolve my main concern (as presented in the first comment of this response).

AuthorsOpenReview ↗

Discussion (1/2)

Thank you for your follow-up engagement in the discussion. We will do our best to address your remaining concerns.

1

The low-quality symmetry information you refer to (whether specified a priori or provided by a symmetry-discovery method) is already discussed under “Approximate symmetry” in Appendix I.1. Specifically, Appendix H.2 tests robustness to symmetry noise, independently of the completeness of the supplied symmetries. Appendices H.3–H.4 also use numerical generators obtained through symmetry discovery, which are approximate.

Regarding symmetry information with an incorrect structural form, Appendix I.1 separately discusses this case under “Misspecified symmetry”. We assess the usefulness of candidate generators by validating the resulting physical equations against observations. Please see our response to point 2 below for the case where no usable symmetry information is available.

Regarding the concern about circular reasoning, symmetry information is a broader structural prior than the governing equation. This information can also be more accessible than the equation itself, since qualitative physical knowledge or transformation relationships in observations may reveal invariances before the detailed dynamics are known. For example, LieNLSD uses sampled points and local differential information to solve the infinitesimal symmetry conditions (Hu et al., 2025, Section 4.2). The local gradients required for this step can be estimated directly from observations using kNN without first fitting a global equation model, as described in Symmetry Discovery for Different Data Types (Hu et al., 2025, Section 4, Algorithm 1). Values and gradients at finitely many sampled points specify local first-order Taylor approximations without uniquely determining a general nonlinear governing equation. This illustrates how symmetry information can be extracted from observations without first identifying the explicit governing equation. The contribution is making this broader prior computationally useful for equation discovery.

2

For the two situations you describe, unusable SD outputs or the absence of useful symmetries, the symmetry-based advantage can indeed be weak or absent. However, our contribution concerns how available symmetry information is used for equation discovery. Limitations of existing symmetry-discovery methods can affect the overall pipeline, but they should not be attributed to our method for using symmetry in equation discovery. Looking ahead, we envision symmetry discovery and symmetry-informed equation discovery progressing together in an upward spiral. As symmetry-discovery methods become more accurate and capable, they can supply better symmetry information for a wider range of systems, enabling our pipeline to become increasingly effective and broadly applicable.

Regarding your concern about systems without useful symmetries: symmetry is a fundamental structural property of many physical laws. Such symmetry structure is therefore widespread across scientific systems, reflecting properties of their governing physical laws. This connection between symmetry and physical laws is illustrated by Noether's theorem, which links continuous symmetries of the action to conservation laws in variational systems. In the extreme case where the absence of symmetry reflects a complete lack of stable physical regularities, there is effectively no physical law to discover. This philosophical limit lies beyond the scope of our study.

AuthorsOpenReview ↗

Discussion (2/2)

3

In fact, our framework can be used with neural PDE models, extending its applicability beyond explicit symbolic equation discovery. A neural network can learn dynamics from invariant features and then predict in the original physical variables.

To test this interface, we compared Raw NN (a neural network using raw variables and derivatives) and DI NN (a neural network using differential-invariant coordinates) on the two-field, two-dimensional RD system in Appendix H.3, using matched MLPs, identical observations and training budgets. After learning in invariant coordinates, DI NN maps its neural outputs back to (ut,vt)(u_t,v_t). Both networks learn the full reaction–diffusion dynamics from observations.

Results are mean ± std over 10 paired initializations on the two held-out test trajectories. RHS and rollout errors are relative L2L^2 errors for the complete physical RHS and two-field state, respectively. Rollouts use the original 100×100100\times100 grid over T=0.2T=0.2. Since RD has field-rotation symmetry, we also evaluate prediction on trajectories transformed by (u,v)↦(−v,u)(u,v)\mapsto(-v,u).

MethodRHS error (↓)Rollout error (↓)Rotated rollout error (↓)
Raw NN(6.70±0.401)×10−3(6.70\pm0.401)\times10^{-3}(1.03±0.117)×10−3(1.03\pm0.117)\times10^{-3}(1.01±0.0886)×10−3(1.01\pm0.0886)\times10^{-3}
DI NN(1.59±0.113)×10−3(1.59\pm0.113)\times10^{-3}(3.31±0.304)×10−4(3.31\pm0.304)\times10^{-4}(3.31±0.304)×10−4(3.31\pm0.304)\times10^{-4}

DI NN reduces mean RHS and trajectory errors by 4.20 and 3.11 times, respectively, with consistent improvements after field rotation. Our framework therefore supports both symbolic equation discovery and neural PDE learning, broadening its applicability to scientific and engineering modeling.

4

Thank you for your positive assessment of the high-dimensional experiment.

5

To test numerical feasibility and stability, we learned field-dependent neural generators from Burgers observations in nonlinear coordinates w=log⁡(1+1.5u)w=\log(1+1.5u). The generators remain in neural form as we compute their prolongations by automatic differentiation and obtain local differential invariants by numerical characteristic integration.

On 32 independent test trajectories, we compare the resulting numerical neural DI with analytic DI and Raw regression, using the same observations and physical-validation selection rule. The table reports the complete physical-RHS relative L2L^2 error. Neural results are mean ± std over 10 generator initializations.

MethodCandidates (↓)Physical RHS error (↓)
Raw regression10106.05×10−26.05\times10^{-2}
Analytic DI664.88×10−44.88\times10^{-4}
Numerical neural DI66(2.35±0.364)×10−3(2.35\pm0.364)\times10^{-3}

Numerical neural DI substantially reduces the mean error compared with Raw regression. To address numerical stability directly, we also doubled the integration steps from 16 to 32 on 512 fixed test points per run. The resulting change was only (5.64±0.842)×10−10(5.64\pm0.842)\times10^{-10} relative to the typical magnitude (RMS) of each invariant, showing negligible sensitivity to integration resolution. These results demonstrate both numerical stability and useful downstream prediction with learned neural generators.

6

Thank you for your positive assessment of the LieGAN experiment. Please see our response to point 1 above for the remaining concern about symmetry quality.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.