acceptodds
Under review as a conference paper at ICLR 2027

Axisymmetry-informed Neural Operators

Abstract

Spherical neural operators have advanced operator learning on the sphere by exploiting group symmetries, with equivariance and invariance providing powerful inductive biases. However, such transformation laws are limited in characterizing the diverse physical responses. In this work, we explore a complementary principle: directly constructing neural operators from known physical response structures. We instantiate this principle through axisymmetry, a common physical structure in which system responses are organized around a distinguished axis. Starting from an axisymmetric Green's function, we derive its corresponding operator solution, which admits a structured spectral convolution characterized by a preferred axis and Legendre polynomials while retaining grid-invariance. This leads to the proposed Axisymmetry-informed Neural Operator (AsiNO). Across spherical physical dynamics, geophysical prediction and cortical modeling, AsiNO achieves superior operator learning, cross-scale generalization and spherical modeling with substantially improved parameter efficiency. These results demonstrate that physical response structure, beyond group symmetry, provides an effective principle for constructing neural operators on the sphere.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operator, Green function, Spherical harmonics, Computational physics

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_rdLGOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a spherical neural operator whose spectral convolution is derived from a physics-motivated assumption: that the system's Green's response depends only on the angle between the output location and a preferred axis. This is a departure from existing spherical neural operator, e.g., SFNO that is equivariant under SO3, which AsiNO claims is too strong a constraint and that the system is symmetric only under rotation around a special direction (i.e., AsiNO offers a less restrictive inductive bias). This choice is physically motivated. The proposed scheme is evaluated over spherical shallow water equations (SWE), WeatherBench, and cortical-surface parcellation (NAMIC, Mindboggle) and AsiNO achieves lowest error on most horizons with far fewer parameters.

Critical strengths
  • AsiNO imposes a prior on Green's function. This is a novel contribution over prior art that impose symmetries over transformations.
  • The results are presented on a diverse selection of benchmarks: simulated SWE, real atmospheric reanalysis (WeatherBench), and biomedical surface parcellation.
  • Zero-shot transfer results shown in Table 2 are impressive. These suggest that asymmetric spectral form is grid-invariant.
  • AsiNO uses far fewer parameters than existing spherical neural operators (SFNO, GSNO).
Critical weaknesses
  • The paper makes no effort of analyzing the learned axes. What directions are actually learned? Do these drift during training?
  • The paper will benefit from ablation where SFNO is also restricted to using the same number of parameters as AsiNO (say by reducing the number of channels for SFNO). Without these comparisons, it is not immediately obvious whether AsiNO wins due to inductive bias or because AsiNO forces fewer parameters.
  • Each channels learns its own axis (right?). If so, it weakens the inductive bias claim.
  • The GA convolution learns two components: \alpha_l and d_c. The reletive contribution of each is not studied. What if d_c is fixed across channels? What if \alpha_l is held constant?
Questions for authors
  • What axes d_c do the channels converge to? For WeatherBench, do they cluster near geographic poles, equator, or distributed uniformaly.
  • Is it possible to see if AsiNO maintains its performance gains over SFNO when SFNO and AsiNO are parameter matched.
1 comment

We sincerely thank the reviewer for the encouraging and constructive feedback, especially the novelty of our Green's response-based formulation, the diversity of our experimental evaluations, and the promising zero-shot transfer results. We greatly appreciate the insightful suggestions regarding interpretability, parameter-matched comparisons, and operator-level ablations.

Weakness 1 and Question 1: Interpretability of Learned Axes.

We sincerely thank the reviewer for this insightful suggestion. We agree that examining the learned axes can provide further insight into the proposed axis-conditioned responses. We would like to clarify that the learnable axes d\mathbf{d} are introduced to parameterize latent axis-conditioned response kernels and structure the spectral convolution, rather than directly representing the physical axes of the underlying systems. We have added a preliminary analysis in the revised appendix, with the new content highlighted in red. The learned polar angles are concentrated near the equatorial plane on WeatherBench (90.09∘±4.48∘90.09^\circ \pm 4.48^\circ). Notably, WeatherBench exhibits a broader distribution and more pronounced variations across operator blocks. We further analyze the training dynamics on SWE, observing modest directional adjustments during training.

Importantly, these learned directions parameterize latent axis-conditioned response kernels rather than directly representing the physical axes of the underlying systems. While these latent directions may capture meaningful directional structures, exploring their connections to physical mechanisms in Earth and brain sciences remains an important direction for future work. The present study primarily focuses on validating the effectiveness of physical-response-informed operator learning.

Weakness 2, 3, 4 and Question 2: Controlled Operator-Level Comparisons and Ablations.

We sincerely thank the reviewer for these constructive suggestions. We agree that parameter-matched comparisons and component-level ablations are important for assessing the contribution of the proposed operator formulation.

To address these concerns, we have conducted additional controlled experiments on SWE. All models follow the same non-spectral network design, training settings, and evaluation protocol, with SFNO/GSNO channel widths adjusted to examine comparable parameter budgets. The results are summarized below, and the full H/V/D results (mean ± std) have been added to the revised manuscript, highlighted in red.

ModelParams (K)Step 10 ↓Step 20 ↓Step 30 ↓
SFNO (C=16C=16)47.3370.249061.197209.69470
SFNO (C=20C=20)73.8890.351951.397808.28860
SFNO (C=24C=24)106.3290.223880.484460.91205
GSNO (C=8C=8)68.3450.242210.613851.12766
GSNO (C=12C=12)111.3450.214330.387092.60863
GSNO (C=16C=16)160.2330.322072.65970176.800
AsiNO (w/o dcd_c)77.1930.195370.303700.48432
AsiNO (shared dd)77.1970.185350.271710.39679
AsiNO (w/o αℓ\alpha_\ell)66.6970.544230.621970.65590
AsiNO77.4490.175140.240800.33352

Values denote the average MRE across H, V, and D (lower is better).

Parameter-matched comparisons. AsiNO achieves the lowest average MRE. At step 30, AsiNO obtains 0.33352, compared with 8.28860 for SFNO (C=20C=20, 73.889K) and 1.12766 for GSNO (C=8C=8, 68.345K).

Operator-level ablations. Removing the learnable axes dcd_c, sharing one axis across channels within each block, or removing the degree-dependent coefficients αℓ\alpha_\ell increases the step-30 average MRE from 0.33352 to 0.48432, 0.39679, and 0.65590, respectively. These results support the contributions of both directional modulation and degree-dependent spectral learning. Importantly, allowing channel-wise axes does not impose a single global symmetry axis; instead, each channel retains its axis-conditioned response parameterization while allowing flexibility across channels.

We have incorporated these results and the corresponding analysis into the revised manuscript (Sec. 5.5), with all additions highlighted in red. We appreciate the reviewer's suggestions, which have strengthened the validation of the proposed formulation.


We sincerely appreciate the careful evaluation, and constructive feedback, which have strengthened this manuscript. We have incorporated the corresponding analyses and experiments into the revised manuscript, with the changes highlighted in red.

Reviewer_dV8POpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper extends the idea of prior-constrained spherical neural operators (e.g., SFNO and GSNO) with a general theoretical framework for constructing spherical neural operators informed by a known physical-response structure. The authors then derive the operator solution for an axisymmetric response structure, in which the response depends only on the angle between uu and the rotated axis RdR\mathbf d.

They explain why the resulting method is parameter-efficient and report improved results on spherical shallow water equations, weather forecasting, and cortical surface modelling benchmarks. They also demonstrate the method’s cross-scale generalization.

Critical strengths
  • The theory section is well written and easy to follow.

The theory section is written quite well and easy to understand and directly relates to the method section. It made easy for me to understand the core idea and contribution of the paper

  • The method improves performance across all benchmarks while using substantially fewer parameters.
Critical weaknesses
  • Section 4 does not explain what n\mathbf n is. From the appendix, I gather that n\mathbf n is a fixed reference direction and that the axis enters only through μ=n⋅d\mu = \mathbf n\cdot\mathbf d. However, the paper never states how n\mathbf n is chosen or discusses the role of μ\mu.
Questions for authors
  • Your model has far fewer parameters than SFNO (0.077M vs 0.755M on SWE), yet its training speed is nearly identical (33.8 vs 33.9 s/epoch, Table 1). Why does the parameter reduction not translate into faster training?
3 comments

We sincerely thank the reviewer for the positive and encouraging assessment of this work, particularly the theoretical formulation and experimental results. We greatly appreciate the thoughtful questions and constructive suggestions, which have helped us improve the clarity and completeness of the manuscript.


Weakness: Reference direction n and the role of μ\mu.

We thank the reviewer for highlighting the importance of clarifying the reference direction and its role in our formulation. In the Green's function formulation, nn is a fixed reference direction (the north pole, n=(0,0,1)n=(0,0,1)). Each latent channel learns its own axis dcd_c, and their relative orientation is characterized by μc=n⋅dc\mu_c=n·d_c. The Legendre factor Pl(μc)P_l(\mu_c) then modulates the spectral response across harmonic degrees through this shared axis parameter.

Question: Parameter efficiency and training time.

We appreciate this insightful question regarding the relationship between parameter efficiency and computational cost. Although AsiNO substantially reduces the number of learnable spectral parameters, evaluating the Legendre polynomials Pl(μc)P_l(\mu_c) and their gradients introduces additional computational operations. Moreover, the SHT/ISHT, MLPs, and other shared network components remain unchanged. Consequently, the parameter reduction does not necessarily translate into a proportional reduction in training time.


We have incorporated additional explanations and analyses addressing both concerns in the revised manuscript (Rebuttal Revision), highlighted in red. These additions clarify the role of the reference direction and axis-conditioned spectral formulation, while providing a more complete discussion of computational efficiency. We sincerely thank the reviewer again for the valuable suggestions, which has strengthened this paper.

Thanks for the response and clarifying my concerns and questions. I would decide on my final score after talking with other reviewers and AC.

We sincerely thank the reviewer for considering our responses. We sincerely appreciate your time and constructive feedback, and would be happy to address any further questions.

Reviewer_1uSUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces a layer based on a filter that is invariant under 3D rotations that leave an axis fixed. Such symmetry is know as axisymmetry and the existence of such symmetry for a physical system with unknown axis of symmetry is the apriori the paper starts with. These layers can be stacked to form neural networks that preserve such symmetry for axis that are learned from the signal (one per channel). They demonstrate in particular superiority of their method on a weather forecasting benchmark.

Critical strengths

A layer based on a filter that is invariant under 3D rotations that leave an axis fixed and therefore to a subgroup of given group. Learning such axis for physical systems where an axisymmetry principle can be applied. An extension of geometric deep learning on the sphere for "mixtures of symmetries", each axis defines a subgroup and encoded as one axis per channel.

Critical weaknesses

None

Questions for authors

None

1 comment

We sincerely thank the reviewer for the positive and encouraging assessment of our work. We particularly appreciate your recognition of the axisymmetry-informed formulation and its extension to channel-wise learnable responses.

We have updated the revised manuscript (Rebuttal Revision), particularly with enhanced presentation and validation, to further strengthen this work (with the changes highlighted in red). We sincerely appreciate your time and thoughtful evaluation, and would be happy to address any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Revisions

Dear Reviewers and AC,

As the discussion period approaches its conclusion, we would like to sincerely thank you for your time, and constructive feedback.

We have carefully considered and responded to the reviewers' comments and updated the manuscript accordingly, with all revisions highlighted in red. We summarize the main improvements below:

  1. Theoretical clarification: Clarified the roles of the reference direction nn and the axis-dependent parameter μ\mu.
  2. Efficiency analysis: Expanded the discussion of parameter efficiency and computational costs.
  3. Learned-axis analysis: Investigated learned-axis distributions and training dynamics on SWE and WeatherBench.
  4. Controlled comparisons: Added parameter-matched comparisons with SFNO and GSNO under consistent settings.
  5. Operator-level ablations: Evaluated the contributions of different operator components.

We sincerely appreciate your valuable feedback and constructive suggestions, which have helped us strengthen this work. We respectfully hope that this summary, together with our responses and revised manuscript, will be helpful for the subsequent discussions and evaluation.

We would be grateful for any further feedback during the remaining discussion period.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.