acceptodds
Under review as a conference paper at ICLR 2027

LCG-SFNO: Low-Rank Clebsch–Gordan Corrections for Global Atmospheric Forecasting on the Sphere

Abstract

Spherical-harmonic neural operators provide an efficient representation for global forecasting, but their discretized pointwise nonlinearities can break rotational equiv- ariance through aliasing. Clebsch–Gordan (CG) tensor products provide exactly equivariant interactions between spherical-harmonic representations, but their com- putational cost grows rapidly with spectral bandwidth and has limited their use at forecasting scale. We introduce LCG-SFNO, which combines low-rank CG interactions within a controlled low-degree band with coefficient-space residual propagation and an exactly projected quadratic nonlinearity, preserving SO(3) equivariance throughout the learned operator. To make the CG interaction practi- cal, we further develop SSHCG, a sparse GPU implementation with a factorized alternative for wider coupled bands. On spherical shallow-water dynamics and the ERA5-based WeatherBench benchmark, LCG-SFNO reduces rotational error to numerical precision—more than three orders of magnitude below SFNO on shallow water and five orders of magnitude below it on WeatherBench—while matching SFNO on shallow-water forecasting and substantially improving fore- cast skill relative to SFNO on WeatherBench. Matched ablations show that CG coupling accelerates convergence, while SSHCG keeps the coupling to about 10% of a training step at the operating bandwidth, making exact equivariant coupling practical at forecasting scale

Primary area:
learning on time series and dynamical systems
Keywords:
Clebsch–Gordan tensor product, Spherical Fourier Neural Operator, SO(3) equivariance

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_PC4nOpenReview ↗
Rating
3
Confidence
1(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To address the limitation that traditional SFNOs cannot strictly satisfy rotational equivariance, this paper proposes LCG-SFNO. It combines low-rank Clebsch-Gordan (CG) cross-order coupling with quadratic nonlinear operations based on exact spectral projection, achieving end-to-end SO(3) equivariance. To reduce computational overhead, the authors further introduce SSHCG, which accelerates CG operations through sparse computation and low-rank factorization. In experiments involving spherical shallow-water equations and the WeatherBench benchmark, this method reduces equivariance error to the level of numerical precision, while maintaining or improving predictive performance compared to standard SFNOs.

Critical strengths
  1. The paper provides a structurally complete and well-validated solution for end-to-end rotational equivariance. To achieve end-to-end SO(3) rotational equivariance, the paper uses low-rank CG coupling to implement strictly equivariant cross-degree nonlinear interactions. It further applies quadratic nonlinear activation on an oversampled grid and removes aliasing errors through exact spectral projection, preventing discrete nonlinear operations from breaking equivariance.

  2. The efficient CG implementation provides clear engineering value. By optimizing CG computation through sparse traversal, output tiling, and fused reduction, the proposed method reduces unnecessary computation and intermediate tensor overhead. Combined with low-rank channel compression and restricted spectral bandwidth design, the CG coupling can be efficiently applied to large-scale spherical forecasting.

Critical weaknesses
  1. The improvement in forecasting performance cannot be fully attributed to CG in the final model. Table 2 shows that strengthening the backbone alone improves one-day ACC from 84.0 to 93.4, while adding CG further improves it to 93.7. However, the final precisely dealiased model achieves 93.3 ACC / 82.1 MSE, which does not show clear advantages compared with the enhanced no-CG backbone with a different nonlinear configuration (93.4 / 81.5). In addition, the SFNO baseline uses spectral bandwidth L=15L=15, while LCG-SFNO uses L=31L=31, and the codec designs are also different. Adding a fully matched final no-CG ablation would significantly strengthen the attribution of the contribution.

  2. Treating all atmospheric variables as scalar spherical fields limits the scope of the claimed physical equivariance. The paper represents all input variables as independent spherical harmonic coefficient channels and adopts a unified SO(3) scalar representation framework. This representation is reasonable for scalar variables such as temperature; however, variables such as zonal and meridional wind are intrinsically tangent vector fields on the sphere rather than scalar fields. Under true three-dimensional rotations, the local basis of tangent vector fields also changes with position, and therefore the two horizontal wind components typically mix during rotation. If wind components are treated as independent scalar channels, the model may guarantee SO(3) equivariance under the scalar-channel representation, but this does not necessarily correspond to the physical rotational transformation law of real atmospheric vector fields. The authors should clarify the current representation scheme and the scope of their equivariance claim.

Questions for authors
  1. Regarding the independent contribution of CG coupling to the final forecasting performance, could the authors provide a no-CG ablation experiment with exactly the same configuration as the final LCG-SFNO model? The current Table 2 B→C ablation demonstrates that CG coupling improves WeatherBench performance under the GELU activation and data-grid nonlinear setting (ACC improves from 93.4 to 93.7, and MSE decreases from 81.5 to 77.9). However, the final model further introduces quadratic activation, exact spectral projection, and different oversampling configurations. Since these designs change the nonlinear pathway and numerical properties of the network, it remains unclear whether the independent forecasting contribution of CG still exists in the final strictly equivariant model. A matched experiment with only CG removed under the final configuration would further validate the contribution of CG to the final model performance.

  2. The paper represents all input variables as independent spherical harmonic coefficient channels. Could the authors clarify the rotational representation of vector-valued atmospheric variables (e.g., horizontal wind fields)? Specifically, are zonal and meridional wind components treated as independent scalar spherical fields, or are they represented using vector field representations that follow the SO(3) transformation law? This point is important for understanding the scope of the paper's claimed "rotational equivariance", because equivariance under a scalar-channel representation is not necessarily equivalent to the physical rotational transformation law of atmospheric vector fields.

  3. Regarding the algorithmic novelty of SSHCG, could the authors further clarify its differences from existing efficient CG computation methods (e.g., Gaunt tensor products and related tensor factorization methods)? While the paper provides comparisons with e3nn and cuEquivariance and relates the factorized route to existing Gaunt-based formulations, the precise algorithmic novelty of SSHCG remains somewhat unclear. Could the authors distinguish the genuinely novel components of the sparse direct kernel from implementation-level optimizations and clarify the additional contribution of the factorized route relative to prior work?

5 comments

Thank you PC4n for the detailed review and constructive suggestions. We address the independent contribution of CG under the final architecture, clarify the rotational representation of atmospheric vector fields, and distinguish the algorithmic contributions of SSHCG from existing efficient tensor-product implementations.

W1 / Q1. Independent Contribution of CG Under the Final Architecture

To isolate the contribution of CG coupling under the final strictly equivariant architecture, we conducted additional matched experiments comparing LCG-SFNO with a no-CG control and a parameter-matched quadratic-branch control. All three models share the same spectral backbone, codec, exactly projected nonlinear pathway, oversampling, and training protocol. The no-CG control removes only the CG branch, while the quadratic-branch control replaces learned CG with an exactly projected quadratic interaction at comparable parameter capacity.

Table A. WeatherBench one-day test performance over three training seeds.

ModelParamsACC (%) ↑MSE (×10−3\times10^{-3}) ↓Equivariance error
No-CG control5.07M93.057 ± 0.02585.230 ± 0.3556.9×10−76.9\times10^{-7}
Quadratic-branch control5.26M93.093 ± 0.03784.725 ± 0.3541.0×10−61.0\times10^{-6}
LCG-SFNO (learned CG)5.24M93.3 ± 0.182.1 ± 1.46.6×10−76.6\times10^{-7}

All three models maintain equivariance errors near numerical precision. LCG-SFNO achieves the lowest forecasting MSE, improving upon both the no-CG control and the parameter-matched quadratic control. This establishes that learned CG coupling provides additional predictive benefits beyond the strengthened backbone, exact quadratic interactions, and increased parameter capacity. The improvement is consistent with the original GELU-based ablation (Table 2, B→C), where adding CG reduces MSE from 81.5 to 77.9.

We also compare the wall-clock time required to reach common validation wRMSE thresholds under the final architecture.

Table B. Training-step cost and wall-clock convergence (seed 42).

ModelTraining stepwRMSE ≤0.265\leq 0.265wRMSE ≤0.2583\leq 0.2583
No-CG control357 ms4.3 hNot reached
Quadratic-branch control398 ms3.9 hNot reached
LCG-SFNO (learned CG)397 ms2.7 h5.0 h

Despite approximately 11% greater per-step cost than the no-CG control, LCG-SFNO reaches the same validation target approximately 37% faster in wall-clock time. Its computational cost is also nearly identical to the parameter-matched quadratic control, while achieving faster convergence and better forecasting accuracy.

Regarding the original SFNO comparison, the backbone improvements and CG contribution are evaluated separately. In Table 2, row A represents the reference SFNO with spectral bandwidth L=15L=15, while row B uses a strengthened backbone with L=31L=31, an improved codec, and regularization. The improvement from A to B therefore reflects these backbone changes rather than CG coupling. The original B→C comparison isolates the contribution of CG under GELU activation, while the new matched experiments establish its independent contribution under the final exactly projected quadratic architecture.

Together, these results distinguish the contributions of the stronger spectral backbone, learned CG interactions, and equivariance-preserving nonlinear pathway. CG consistently improves forecasting accuracy within matched configurations and provides substantial wall-clock convergence benefits under the final architecture.

W2 / Q2. Representation of Wind Fields and Scope of Equivariance

LCG-SFNO represents each input channel as a scalar field whose spherical-harmonic coefficients transform through Wigner D-matrices. Accordingly, the joint equivariance relation in Eq. (3) and the measurements in Table 1 refer to simultaneous rotations of the scalar channels and spatial conditioning fields.

This representation directly corresponds to physical rotations for scalar quantities, including temperature, height, vorticity, and divergence. In the WeatherBench experiments, zonal and meridional winds are treated as separate scalar channels, following the benchmark configuration. Under rotations about Earth's polar axis, this representation is compatible with the physical transformation of horizontal wind components. Under general SO(3) rotations, however, the local tangent basis also transforms, causing the wind components to mix. Thus, the reported WeatherBench equivariance is defined under the specified scalar-channel representation.

The architecture also supports physical vector-field equivariance through a Helmholtz–Hodge representation. A tangent velocity field can be represented by scalar vorticity and divergence using vector spherical harmonics. These scalar fields are processed by the same LCG-SFNO spectral blocks and subsequently reconstructed into the velocity field. Since the transformations between vector fields and their scalar potentials commute with rotations, this formulation preserves the physical SO(3) transformation law without modifying the learned CG coupling or exactly projected nonlinear pathway.

This representation is used in our SSWE experiments, where height, vorticity, and divergence constitute the dynamical state. The corresponding wind-field interface is also implemented, while WeatherBench retains the component-wise representation for consistency with the evaluated baselines.

The distinction between these representations clarifies the scope of the results. The SSWE experiments establish equivariance under the physical rotational representation, while the WeatherBench experiments demonstrate numerical equivariance under the specified scalar-channel action. This distinction will be made explicit alongside Eq. (3) and in the appendix.

Q3. Algorithmic Contributions of SSHCG Relative to Existing Efficient CG Methods

SSHCG is designed for broad-band Clebsch–Gordan coupling in spherical spectral operators, where interactions span all degrees up to LcL_c, outputs extend to 2Lc2L_c, and both parity sectors are included. This setting differs from the predominantly low-angular-degree workloads of equivariant interatomic models addressed by implementations such as OpenEquivariance [1] and FlashTP [2]. Existing work also exploits CG sparsity and efficient Gaunt products [3, 4]. SSHCG focuses on efficient full-spectrum coupling while preserving learnable degree-path interactions.

The implementation provides two complementary computational routes.

Sparse direct route. This route evaluates the original CG contraction with independent learnable weights for each admissible degree path. Its main algorithmic features are the reuse of channel-independent angular sparsity patterns and direct accumulation of multiple degree-path contributions into output tiles. These operations avoid materializing large intermediate tensors while retaining the unrestricted degree-path parameterization.

Factorized route. This route applies a CP factorization to the learnable degree-path weights. It combines efficient Gaunt products for even-parity interactions with spherical Poisson brackets for odd-parity interactions, supporting both parity sectors within a unified learnable spectral operator. The factorized route reproduces the direct contraction for weights represented by its factorization, while the sparse direct route retains independent degree-path weights.

These two routes provide complementary tradeoffs between parameterization flexibility and computational efficiency. The sparse direct route preserves the original learnable CG contraction, while the factorized route improves scalability for wider coupling bandwidths.

Table C. Forward-and-backward evaluation time of a single CG block (Table 3).

Coupling bandwidthDenseSparse directFactorized
12/2416.6 ms4.6 ms4.1 ms
31/31855 ms35.6 ms5.3 ms

At 31/31 bandwidth, the sparse direct and factorized implementations achieve approximately 24-fold and 161-fold speedups over the dense reference, respectively. The factorized route also exhibits weak runtime growth across the tested bandwidths, increasing from 3.2 ms at 4/8 to 5.3 ms at 31/31.

The default LCG-SFNO configuration uses the sparse direct implementation at 12/24 bandwidth to retain independent learned degree-path interactions. The factorized route provides an efficient alternative for wider coupling bands.

The principal contribution of SSHCG is therefore the combination of full-spectrum CG coupling, learnable degree-path interactions, and computationally efficient sparse or factorized evaluation. The sparse direct route reduces intermediate-memory and contraction costs without restricting the learned degree-path weights, while the factorized route extends efficient spectral products to both parity sectors. Together, they make learned CG interactions practical at forecasting-relevant spectral resolutions.

References

[1] Bharadwaj et al. (2025). An Efficient Sparse Kernel Generator for O(3)-Equivariant Deep Networks. ACDA.

[2] Lee et al. (2025). FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials. ICML.

[3] Xie et al. (2025). The Price of Freedom: Exploring Expressivity and Runtime Tradeoffs in Equivariant Tensor Products. ICML.

[4] Luo et al. (2024). Enabling Efficient Equivariant Operations in the Fourier Basis via Gaunt Tensor Products. ICLR.

Reviewer_MbV7OpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LCG-SFNO, which augments a spherical spectral backbone with low-rank Clebsch–Gordan (CG) coupling within a restricted low-degree band. Coefficient-space residuals and an exactly projected quadratic nonlinearity preserve SO(3) equivariance, while the proposed SSHCG implementation reduces coupling costs. The method achieves near-machine-precision equivariance, matches SFNO on shallow-water forecasting, and improves over the reference SFNO on WeatherBench.

Critical strengths
  1. Careful treatment of end-to-end equivariance. The paper identifies nonlinear aliasing as a limitation of spherical-harmonic architectures and addresses it explicitly. Equivariance errors around (10−710^{-7}) on both benchmarks provide clear numerical support.

  2. Practical implementation improvements. SSHCG reduces per-block forward–backward latency from 16.6 ms to 4.6 ms at the operating bandwidth. Comparisons with existing tensor-product libraries further support its computational utility.

Critical weaknesses
  1. The independent benefit of CG coupling is not compelling enough. In Table 9, strengthening the backbone, codec, and regularization reduces MSE from 187.7 to 81.5; adding CG further reduces it to 77.9. Most of the improvement therefore comes from architectural changes outside the proposed coupling. The paper does not clearly establish what distinctive advantage the restricted-band CG correction offers over simpler equivariant nonlinear alternatives.

  2. The key ablations do not match the final architecture. The incremental ablations use GELU, whereas the final model uses an exactly projected quadratic activation. A matched no-CG control within the final architecture is missing. Consequently, the additional accuracy and optimization benefits of CG on top of an already strictly equivariant backbone remain unclear.

  3. The practical forecasting advantage remains limited. Performance is comparable to SFNO on shallow water, while ClimaX or SHNet leads on most aggregate WeatherBench metrics. Faster convergence is primarily demonstrated in epochs rather than wall-clock time to comparable accuracy, and real-weather evaluation is limited to (32X64) resolution.

Questions for authors
  1. Could you remove CG from the final quadratic-activation architecture while keeping all other settings and model capacity comparable, and report multi-seed forecasting accuracy, equivariance error, and wall-clock time to the same validation target?

  2. The exactly projected quadratic nonlinearity already induces interactions across harmonic degrees. What additional capability does explicit CG coupling provide? Could a budget-matched comparison with a simpler equivariant quadratic branch establish the value of the learned degree-path weights and low-rank CG parameterization?

  3. Could you provide accuracy-versus-wall-clock comparisons against the strengthened no-CG backbone, and evaluate whether the gains persist at higher spatial resolutions or longer forecast horizons?

4 comments

Thank you for the constructive comments and suggestions. We conducted additional matched ablations under the final architecture, including a parameter-matched equivariant quadratic control, and evaluated forecasting accuracy, wall-clock convergence, and long-horizon prediction under different training-data budgets.

W1 / Q2. Why Learned CG Rather Than a Simpler Equivariant Quadratic Interaction?

The exactly projected quadratic activation already introduces cross-degree interactions through pointwise multiplication, corresponding to Gaunt coupling in spherical-harmonic space. However, these interactions follow the fixed Gaunt structure and are restricted to even-parity degree paths. In contrast, the learned CG correction introduces independent trainable weights γℓ1ℓ2ℓ\gamma_{\ell_1\ell_2\ell} for admissible degree paths and includes both even- and odd-parity interactions. This provides a more flexible parameterization of equivariant spectral interactions while the low-rank formulation limits computational cost.

To isolate the value of these learned interactions, we introduce a parameter-matched quadratic-branch control that replaces the CG correction with an exactly projected quadratic interaction. The control retains the degree-wise learnable P/Q/O projections and the same input/output coupling bandwidth (12/24). Its fixed Gaunt interaction covers 819 even-parity paths, compared with 1,469 admissible paths in the learned CG correction. The parameter counts are closely matched (5.26M for the quadratic control and 5.24M for LCG-SFNO).

The matched experiments reported in Tables A and B below show that learned CG coupling achieves lower one-day WeatherBench MSE than the quadratic control (82.1 versus 84.73) and reaches the same validation target in 2.7 rather than 3.9 hours. The CG branch also has lower forward-and-backward evaluation cost in our implementation (5.04 ms versus 5.84 ms), with nearly identical full training-step times.

These results demonstrate that learned CG coupling provides benefits beyond the equivariant interactions already available through quadratic activation. Its adaptive degree-path interactions improve forecasting accuracy and optimization efficiency while preserving numerical rotational equivariance.

W2 / Q1. Independent Contribution of CG Under the Final Architecture

To isolate the contribution of CG under the final strictly equivariant architecture, we conducted matched experiments comparing three models:

  • No-CG control: removes the CG branch while retaining the final backbone, codec, exactly projected quadratic activation, oversampling, and training protocol.
  • Quadratic-branch control: replaces learned CG with an exactly projected quadratic interaction at the same coupled bandwidth and a comparable parameter count.
  • LCG-SFNO: retains the full architecture with learned CG coupling.

All three models therefore share the same equivariance-preserving nonlinear pathway, allowing the independent predictive contribution of learned CG interactions to be evaluated.

Table A. WeatherBench one-day test performance over three training seeds.

ModelParamsACC (%) ↑MSE (×10−3\times10^{-3}) ↓Equivariance error
No-CG control5.07M93.057 ± 0.02585.230 ± 0.3556.9×10−76.9\times10^{-7}
Quadratic-branch control5.26M93.093 ± 0.03784.725 ± 0.3541.0×10−61.0\times10^{-6}
LCG-SFNO (learned CG)5.24M93.3 ± 0.182.1 ± 1.46.6×10−76.6\times10^{-7}

All three models achieve equivariance errors near numerical precision, confirming that the exactly projected nonlinear pathway preserves rotational consistency independently of CG. Within this strictly equivariant setting, LCG-SFNO achieves the lowest forecasting MSE, improving upon both the no-CG and parameter-matched quadratic controls. This result is consistent with the original GELU-based ablation (Table 2, B→C), where introducing CG reduces MSE from 81.5 to 77.9.

We further compare the wall-clock time required to reach common validation wRMSE thresholds, including both training and validation costs.

Table B. Wall-clock convergence under the final architecture (seed 42).

ModelTime/epochwRMSE ≤0.265\leq 0.265wRMSE ≤0.2583\leq 0.2583
No-CG control815 s4.3 h (19 epochs)Not reached
Quadratic-branch control930 s3.9 h (15 epochs)Not reached
LCG-SFNO (learned CG)898 s2.7 h (11 epochs)5.0 h (20 epochs)

Despite approximately 11% greater training-step cost than the no-CG control, LCG-SFNO reaches the same validation threshold approximately 37% faster in wall-clock time. It also reaches the stricter threshold of 0.2583 within the training budget, whereas neither control reaches that target. Compared with the quadratic-branch control, LCG-SFNO has nearly identical full training-step cost while achieving faster convergence and better final accuracy.

These matched experiments establish that CG provides independent forecasting and optimization benefits even when the backbone and nonlinear pathway already enforce rotational equivariance. The improved prediction accuracy arises from the learned spectral interactions rather than from equivariance enforcement or additional parameter capacity alone.

W3 / Q3. Practical Forecasting Performance and Computational Efficiency

We further evaluate the practical benefits of LCG-SFNO through wall-clock convergence and long-horizon forecasting under controlled training-data budgets.

Wall-clock training efficiency. The matched results in Table B demonstrate that the convergence improvement reported in Figure 3 translates into actual computational savings. LCG-SFNO reaches a validation wRMSE of 0.265 in 2.7 hours, compared with 4.3 hours for the strengthened no-CG backbone and 3.9 hours for the parameter-matched quadratic control. CG coupling therefore reduces total training time despite its modest additional per-step cost. The stricter validation target is also reached only by LCG-SFNO within the evaluated training budget.

Long-horizon forecasting and data efficiency. We conducted additional experiments on SSWE with controlled high-frequency spectral components, using 25%, 50%, 75%, and 100% of the available independent training data. All models use the same 800 optimization updates and 12,800 sample accesses, allowing performance to be compared under the same computational and sample-access budgets.

The following table reports relative L2L^2 rollout errors (×10−3\times10^{-3}) over three training seeds, evaluated on 50 independent test initial conditions.

Training dataSFNO 5 hLCG-SFNO 5 hSFNO 10 hLCG-SFNO 10 h
25%68.90 ± 1.9466.63 ± 3.1686.39 ± 2.6685.55 ± 7.32
50%68.08 ± 0.8565.59 ± 1.4182.97 ± 4.8179.24 ± 2.49
75%69.24 ± 5.3167.09 ± 4.3186.35 ± 11.6384.44 ± 9.08
100%67.19 ± 4.7265.84 ± 3.7484.79 ± 7.1083.86 ± 7.19

LCG-SFNO achieves lower mean rollout errors at both horizons across all training-data budgets. The improvement persists when the number of independent training samples is reduced, demonstrating the predictive benefits of the complete architecture under limited data availability.

Spatial resolution and forecast horizons. The original 256×512 SSWE experiments (Tables 1 and 4) already demonstrate LCG-SFNO at substantially higher spatial resolution than the 32×64 WeatherBench evaluation. These results establish competitive long-horizon accuracy alongside substantially improved rotational consistency. On WeatherBench, LCG-SFNO also remains competitive across the evaluated one-, three-, and five-day forecast horizons (Table 5), while the new matched experiments establish the independent contribution of CG at the one-day horizon.

Together, the original experiments and additional evaluations establish three complementary benefits: numerical rotational equivariance through exact spectral projection, improved forecasting accuracy and wall-clock convergence through learned CG coupling, and computationally efficient spectral interactions through SSHCG. The high-resolution SSWE results further support the scalability of the approach beyond the WeatherBench evaluation grid.

Thank you for the additional experiments. The matched no-CG and parameter-matched quadratic controls are useful and address an important part of my concern. However, the scope and robustness of this benefit remain unclear. The matched comparisons are limited to one-day WeatherBench forecasting at 32×6432\times64. The original 256×512256\times512 SSWE results demonstrate operation on a larger spatial grid, but do not establish whether the same CG-specific gains persist at higher resolution or larger coupling bandwidths. The new long-horizon experiments compare the complete architectures, and several improvements are modest relative to the reported variability, leaving the independent long-horizon contribution of CG unresolved.

Your rebuttal strengthens the evidence for the method, but does not yet sufficiently resolve my concerns about the generality and robustness of the claimed benefits. I therefore maintain my original score but lower my confidence.

Thank you for following up and for acknowledging the updates in our rebuttal. To directly test whether the gains from CG coupling hold up at higher resolutions and longer horizons, independent of broader architectural differences, we ran an additional controlled ablation on the 256×512 SSWE benchmark. Here, the No-CG baseline and LCG-SFNO share the exact same backbone, quadratic nonlinearity, dealiasing strategy, and training recipe; the only difference is the CG coupling module itself.

The relative L2L_2 rollout errors (×10−3\times 10^{-3}, mean ±\pm std over five seeds) are shown below:

Model1 h ↓5 h ↓10 h ↓
No-CG56.51 ± 0.38119.37 ± 2.26211.27 ± 7.78
LCG-SFNO (CG)56.90 ± 0.50115.80 ± 0.70199.90 ± 0.90

At the 1 h mark, both variants perform similarly. As the rollout extends, however, the effect of CG coupling becomes much clearer: it lowers the error by 3.0% at 5 h and 5.4% at 10 h, while also noticeably tightening the variance across seeds (dropping from 7.78 to 0.90 at 10 h). Because all other architectural choices are held fixed, this trend confirms that CG coupling actively stabilizes long-horizon rollouts at higher spatial resolutions.

We also agree with your point that scaling to broader CG coupling bandwidths and higher-resolution WeatherBench settings would be a valuable next step. We will explicitly note the scope of our current experiments in the revised manuscript and explore those larger-scale settings in revision with longer time. Thank you again for the constructive discussion.

Reviewer_CNsROpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposed a new Fourier neural operator that have spherical invariance equivariance properties.

It uses special operations between spherical harmonic and polynomial nonlinearities to achieve that.

The paper also discuss a number of experiments

Critical strengths

There are really 2 main problems with the paper

  1. Even though equivariant nets have very nice theoretical properties they seem to perform worse than simple networks with many data augmentations.

  2. The main problem is that for the main application here, coriolis force is not equivariant with respect to latitude. So why use the network here

  3. The results are not convincing. It is not clear to me that any improvement is a result of the net

Critical weaknesses

See above

Questions for authors

See above

2 comments

Thank you for the constructive feedback and suggestions. We address the three weaknesses below, supported by additional experiments on rotational augmentation, latitude-dependent dynamics, and the architectural contributions of LCG-SFNO.

W1. Equivariant Networks vs. Data Augmentation

To directly compare architectural equivariance with data augmentation, we conducted an additional experiment on SSWE, training SFNO with random SO(3) rotations applied jointly to each input–target pair. Rotations were performed in spherical-harmonic coefficient space using Wigner D-matrices to avoid spatial interpolation errors. All other training settings remained unchanged. Results are averaged over five seeds and evaluated on the same 50 fixed test initial conditions.

SSWE relative L2L^2 rollout error (×10−3\times 10^{-3}) and rotational equivariance error.

Model1 h5 h10 hEquivariance error
SFNO57.7 ± 0.8123.4 ± 6.6229.7 ± 20.79.5×10−49.5\times10^{-4}
SFNO + rotation augmentation57.5 ± 0.9121.7 ± 8.3223.7 ± 23.19.2×10−49.2\times10^{-4}
LCG-SFNO56.9 ± 0.5115.8 ± 0.7199.9 ± 0.98.2×10−78.2\times10^{-7}

Rotation augmentation modestly improves SFNO's mean forecasting accuracy but leaves its equivariance error nearly unchanged. In contrast, LCG-SFNO achieves the lowest mean rollout error at all three horizons, with substantially lower variation across training seeds at longer horizons, while reducing equivariance error by more than three orders of magnitude.

From a theoretical perspective, data augmentation encourages equivariance through the training distribution [1, 2], whereas architectural equivariance enforces consistent transformations by construction [3, 4], independently of the orientations observed during training. This structural constraint supports generalization across unseen rotations without requiring the model to learn rotational consistency from additional examples [2, 5]. Our experiments demonstrate that architectural equivariance can be achieved alongside improved forecasting accuracy and long-horizon stability.

W2. Coriolis Force and the Use of an Equivariant Network

The latitude dependence of the Coriolis force is compatible with the conditional rotational equivariance used in LCG-SFNO. With Earth's rotation axis fixed, the Coriolis parameter f=2Ωsin⁡ϕf=2\Omega\sin\phi is invariant under rotations about this axis, corresponding to the subgroup SO(2)⊂SO(3)SO(2)\subset SO(3).

More generally, our architecture supports joint SO(3) equivariance,

F(Rx,Rc)=RF(x,c),F(Rx,Rc)=RF(x,c),

as defined in Eq. (3), where cc denotes spatial conditioning fields. Rotations about Earth's axis leave the Coriolis field unchanged, while general rotations transform the conditioning fields together with the atmospheric state. This formulation accommodates latitude-dependent dynamics through rotation-consistent operations rather than imposing identical behavior at every latitude.

In WeatherBench, latitude, orography, and land–sea mask serve as conditioning fields (Appendix D.1). Latitude encodes the Coriolis parameter, whose spatial dependence has degree-one spherical-harmonic structure. The dealiased quadratic pathway and low-degree CG coupling enable interactions between atmospheric states and these conditioning fields, allowing the model to represent spatially varying dynamics.

The experimental results support this formulation. Adding positional embeddings to SFNO provides no forecasting improvement (Appendix E.4), while LCG-SFNO achieves the highest one-day ACC among the evaluated models (Table 5) and numerical equivariance under joint rotations of the state and conditioning fields (Table 1). These results establish that rotation-consistent spectral operations can accurately model latitude-dependent atmospheric dynamics under the specified field representation.

W3. Improvements and Their Architectural Origins

The contributions of LCG-SFNO extend beyond forecasting accuracy to training efficiency, rotational consistency, and generalization. We evaluate these aspects through complementary metrics and controlled comparisons, with ablations identifying the contributions of individual architectural components.

Forecasting accuracy and stability. On WeatherBench, LCG-SFNO substantially improves upon the reference SFNO across all three forecast horizons. At one day, ACC increases from 84.0 to 93.3, while MSE decreases from 187.7 to 82.1 (Table 5). On SSWE, LCG-SFNO achieves the lowest mean rollout error at one, five, and ten hours among the five evaluated models, with the fewest parameters (Table 4).

Training efficiency. Figure 3 and Table 2 demonstrate faster convergence toward comparable validation errors. Additional matched experiments under the final architecture confirm this benefit: LCG-SFNO reaches the same validation-error threshold in 2.7 hours, compared with approximately 4.3 hours for the no-CG control, reducing wall-clock training time by approximately 37%. The matched architecture and timing results are detailed in our response to Reviewer MbV7 (W2/Q1).

Rotational consistency. LCG-SFNO enforces rotational equivariance by construction, reducing equivariance error by more than three orders of magnitude on SSWE and approximately five orders on WeatherBench, reaching numerical precision (Table 1). Under the specified spherical-field representation, this property holds for arbitrary SO(3) rotations, including orientations unseen during training.

Generalization across unseen orientations. To evaluate whether architectural equivariance translates into stable forecasting accuracy, we conducted an additional WeatherBench experiment by shifting the prime meridian. All input fields, spatial conditioning fields, and targets were shifted jointly by seven longitude offsets (45°–315°), without interpolation, retraining, or augmentation. These shifts belong to the Coriolis-compatible SO(2) subgroup discussed in W2. Results are averaged over three training seeds.

WeatherBench one-day ACC (%) and MSE (×10−3\times10^{-3}) under original and shifted orientations.

ModelACC originalACC shiftedMSE originalMSE shifted
ClimaX93.2 ± 0.381.3 ± 0.276.8 ± 0.5195.8 ± 2.2
FourCastNet92.8 ± 0.186.4 ± 0.182.8 ± 0.3155.1 ± 0.8
SHNet92.4 ± 0.085.4 ± 0.192.7 ± 0.1171.0 ± 1.4
SFNO84.0 ± 0.084.0 ± 0.0187.7 ± 0.2187.7 ± 0.2
LCG-SFNO93.3 ± 0.193.3 ± 0.182.1 ± 1.482.1 ± 1.4

LCG-SFNO maintains its forecasting accuracy across all tested orientations, whereas ClimaX, FourCastNet, and SHNet lose 6.4–11.9 ACC points. SFNO also preserves its accuracy under longitude shifts, but at substantially lower forecast skill. LCG-SFNO therefore combines high predictive accuracy with orientation-independent forecasting under this physically relevant rotation subgroup, without requiring rotational augmentation.

The ablation results further distinguish the contributions of individual components. Strengthening the full-band spectral backbone, codec, and regularization substantially improves forecasting accuracy over the reference SFNO (Table 2, A→B). Learned CG coupling provides additional improvements in accuracy and convergence, as confirmed by matched experiments under both nonlinear configurations (Table 2, B→C; Reviewer MbV7, W1/Q2 and W2/Q1). The exactly projected quadratic pathway eliminates nonlinear aliasing in the retained spectrum and enables numerical SO(3) equivariance, while SSHCG makes CG coupling computationally practical (Table 3; Reviewer PC4n, Q3).

Together, these results establish the complementary roles of the stronger spectral backbone, learned CG coupling, and exactly projected nonlinear pathway. The additional orientation-shift experiment further demonstrates that the resulting architecture combines strong forecasting skill with stable performance across physically relevant rotations.

References

[1] Gerken et al. (2022). Equivariance versus Augmentation for Spherical Images. ICML.

[2] Wang et al. (2022). Data Augmentation vs. Equivariant Networks: A Theory of Generalization on Dynamics Forecasting. arXiv:2206.09450.

[3] Cohen & Welling (2016). Group Equivariant Convolutional Networks. ICML.

[4] Kondor & Trivedi (2018). On the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact Groups. ICML.

[5] Elesedy & Zaidi (2021). Provably Strict Generalisation Benefit for Equivariant Models. ICML.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.