acceptodds
Under review as a conference paper at ICLR 2027

Equivariant DeepONets for Symmetry-Aware Operator Learning

Abstract

Neural operators constitute a powerful class of deep learning architectures within scientific machine learning which learn mappings between function spaces, yielding natural and effective surrogates for parametric systems of partial differential equations (PDEs). A classical and potent method is the Deep Operator Network (DeepONet), which separates an input-dependent branch and output-coordinate trunk before fusing. Simultaneously, equivariant deep learning has been shown to be a powerful mechanism to imbue networks with knowledge of physical symmetries. In this work, we construct equivariant DeepONets through a representation-theoretic framework in which the observation, branch, trunk, and fusion maps obey compatible group actions. Our construction uses equivariant linear layers and symmetry-compatible branch-trunk fusion. We show that finite observations are compatible with equivariance when they recover a finite-dimensional invariant input representation or when the sensor family is closed under the spatial group and prove a universal approximation theorem for continuous equivariant operators by equivariant DeepONets. We evaluate the framework on numerous PDE systems with different symmetry groups, equivariance actions, and system sizes, including Allen–Cahn dynamics on the sphere, non-rotating spherical shallow water equations, thermoelasticity on an annulus, and a two-component Gross–Pitaevskii system. The equivariant architectures achieve order-of-magnitude reductions in prediction error and equivariance diagnostics over baseline DeepONets while often simultaneously possessing fewer learnable parameters, and the finite-sensor variants achieve comparable results. We also compare favorably to both appropriate FNO architectures and DeepONets with pre-chosen symmetry-aware trunks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific machine learning, neural operators, geometric deep learning, partial differential equations

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_6H2dOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper makes DeepONets exactly equivariant by requiring all of their parts (observation, branch, trunk, and fusion) to respect the same symmetry. It also gives conditions under which finite sensor measurements keep this property. On four PDE problems with different symmetries, the equivariant DeepONets achieve lower errors than baseline DeepONets and FNO references, usually with fewer parameters.

Critical strengths

S1. The framework is clean and general. Treating equivariance as a property of the whole DeepONet factorization, rather than of the trunk alone, is a useful point, and the paper backs it up with both theory and an ablation (the symmetry-aware trunk model).

S2. The experiments cover four quite different symmetry settings, including mixed scalar and vector fields, reflections, and an internal U(2) symmetry. The equivariant models behave consistently across all of them, and the authors are open about limitations such as the larger model needed for Gross–Pitaevskii.

Critical weaknesses

W1. I am not fully convinced the comparisons are fair.

W1.1 The baselines look under-trained. The vanilla DeepONet has one-step relative L2 errors of 0.34 (Allen–Cahn), 0.89 (shallow water), and 0.32 (Gross–Pitaevskii), and its five-step error on shallow water is around 1.0. The symmetry-aware trunk model uses 24.1M parameters on thermoelasticity (vs. about 1M for the proposed model) yet does no better than the vanilla DeepONet. I suspect the "order-of-magnitude" improvements partly reflect weak baselines.

W1.2 It is unclear whether all models see the same input. Section 4 contrasts "a conventional sparse DeepONet" with "a full-observation equivariant model," but the main text does not say how many sensors each model gets. Part of the gap may come from the input rather than from equivariance.

W2. Some natural baselines are missing.

W2.1 There is no comparison with simpler ways of adding symmetry, such as data augmentation or canonicalization (e.g., Shumaylov et al., 2025, cited in the paper). These are needed to show that the full representation-theoretic construction is worth its complexity.

W2.2 Recent group-equivariant neural operators such as G-FNO (Helwig et al., 2023) are not compared, although G-FNO could be applied at least to the Gross–Pitaevskii benchmark on the square.

W3. Efficiency is argued only through parameter counts. Tensor-product nonlinearities are usually expensive, so training time, inference time, and memory should also be reported.

Questions for authors

Q1. (W1.1) How were the baseline DeepONets tuned? It would help to see the tuning budget and training curves. In particular, why does the 24.1M-parameter symmetry-aware trunk model not improve over the vanilla DeepONet?

Q2. (W1.2) How many sensors does each model receive, and where are they placed? Could you add a vanilla DeepONet that gets exactly the same input as the equivariant model (e.g., the same reconstructed coefficients)? That would isolate the effect of equivariance.

Q3. (W2) Could you include (a) a standard DeepONet trained with symmetry augmentation, (b) a canonicalization baseline, and (c) G-FNO where the symmetry group allows it?

Q4. (W3) Could you report training time, inference time, and memory for all models?

Reviewer_aXCROpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces Equivariant DeepONets for operator learning, which ensures equivariance by applying compatible group actions to observation, branch, trunk and fusion. Finite-sensor equivariance conditions and universal approximation theorems are established. Experiments have verified the effectiveness and efficiency of this method on multiple PDE systems. However, there are still several key issues.

Critical strengths

This paper systematically establishes the equivariance compatibility conditions for observation, branch, trunk, and fusion, and proves the equivariance of the overall architecture. Experiments on multiple PDE systems have verified the effectiveness of this method and demonstrated that it typically requires fewer parameters.

Critical weaknesses

The current manuscript lacks ablation experiments to independently validate the contribution of equivariance constraints under the same input representations, making it difficult to clearly identify the main source of performance improvement. The current benchmarks all satisfy pre-defined exact symmetry, making it difficult to determine the applicability of the proposed equivariance constraints in more general physical systems. The paper lacks verification of noisy/sparse observations, approximate symmetry, and symmetry-breaking perturbations. There is a lack of multi-seed stability, long-term prediction, and comparisons with other SOTA (recent FNO/SFNO variants).

Questions for authors

Does the method presented in this paper require specific architectural modifications for different symmetry groups and PDE systems, or can a unified model architecture be adopted? When the physical system only satisfies approximate equivariance rather than exact symmetry, will the methods presented in this paper still be applicable?

Reviewer_Y5uuOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes an equivariant version of DeepONet by making the branch and trunk nets both equivariant. The dot product is also extended to tensor product.

Critical strengths
  • The paper presented a natural extension for considering equivariance in DeepONets.
Critical weaknesses
  • The baseline error is very large. Some possible reasons: The vanilla DeepONet gets only 1024 point sensors while the equivariant model gets the full field. So part of the gap could be information instead of symmetry. Also, it seems that the equivariant model predicts every harmonic coefficient used to generate the data without further cutoff.

  • Both the introduction and experiment sections only cover DeepONet and FNO, while a significant number of UNet-based and (Vision)-Transformer-based models are developed in recent years. Also the models are only tested on self-generated datasets. There is no benchmark on existing public datasets.

  • A reasonably designed baseline trained with symmetry augmentation can be helpful.

Questions for authors
  • Why is the finite sensing important to consider? Can you give an example of case 1 sampling?
  • The extension to equivariance seems to be straightforward. What are the most surprising elements when considering such an extension?
  • Are the baselines well designed and tuned?
  • Do the proof or experiments use the general tensor product fusion?
  • Are the FNO baselines equivariant?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.