Pith.

Artificial Intelligence · arXiv:2611.16356 v1 · submitted 12 Nov 2026 · 67 references

DIAGNO: Diagonal Spherical Neural Operators for Heterogeneous Earth Dynamics Modeling

Herui Li, Bin Lu, Haonan Qi, Lei Zhou, Luoyi Fu, Xinbing Wang, Meng Jin

The paper says

“Accurate forecasting of Earth's dynamic systems is vital in geophysics, where neural operators offer a powerful data-driven paradigm. Standard operators often treat the Earth as a flat Euclidean plane, introducing geometric artifacts.”

more of the abstract

In contrast, Spherical Neural Operators (SNOs) resolve this mismatch by operating within the spherical spectral space, naturally adapting to spherical geometries. However, for computational efficiency, existing SNOs typically rely on the rotation equivariance assumption. This assumption simplifies the transitions of physical states into isolated spectral modal evolutions, thereby over-idealizing the Earth as a homogeneous system. To solve this problem, we no longer rely on the rotation equivariance assumption that isolates spectral modal evolution, but directly derive an explicit modal interaction kernel and propose Diagonal Spherical Neural Operators named DIAGNO. DIAGNO explicitly parameterizes cross-modal interactions through spectral modal decomposition, thereby preserving the inherent heterogeneity of Earth dynamics. Furthermore, leveraging the unique diagonal structure of the spherical spectrum, we design inter- and intra-diagonal interaction mechanisms to natively capture essential zonal and meridional dynamics, respecting intrinsic geophysical regularities. Extensive experiments across simulated fluid dynamics and real-world Earth system reanalysis (including atmosphere and ocean) datasets demonstrate DIAGNO's superior capability, consistently achieving state-of-the-art performance in multi-step forecasting. Beyond this, we validate DIAGNO's practical analytical value in professional geophysical domains by assessing its fidelity in capturing climatological anomalies and maintaining physical consistency.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. DIAGNO is a solid, well-executed paper that introduces a genuinely novel architectural idea for spherical neural operators and backs it with thorough experiments. The missing error bars are a real reporting gap, but the central claim holds up. This line is wrong

4 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 67 references say.

The paper says, at Section 4.1, paragraph following Eq. (9)

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

The physical interpretation of the diagonal index sets appears to invert the standard roles of the spherical-harmonic indices and is internally inconsistent. For Y_l^m, m is the zonal (longitudinal) wavenumber and l - |m| counts the number of meridional (latitudinal) nodes; on a fixed diagonal D_k, the quantity l - m = k - 1 is constant, so the meridional structure is fixed while m grows with l.

more of objection 1

The sentence "modes within the same D_k ... share a unified zonal pattern, where an increasing degree l proportionally raises the longitudinal wavenumber to capture denser meridional oscillations" is contradictory, because raising the longitudinal wavenumber changes the zonal structure, not the meridional one, and the following sentence similarly attributes "progressively denser zonal structures" to increasing k, which raises the latitudinal wavenumber. Since the zonal/meridional distinction is the central physical motivation of the factorization (contribution bullet 2) and is reused in the interpretability analysis of Section 5.4 and Figure 7, the authors should correct this description and re-derive the claimed scale-resolution correspondence. The architecture and numerical results are unaffected by this fix.

The paper says, at Appendix A.2 and Eq. (13)

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

Appendix A.2 overstates what the Best Approximation Theorem establishes for the conjugate-reconstruction design. The minimizer of min_{W_rec} ||u_hat - W_rec W_trans u_hat||_H over linear reconstructions is the Moore-Penrose pseudoinverse W_trans^dagger = (W_trans^H W_trans)^{-1} W_trans^H (for full column rank), which is not in general proportional to the adjoint W_trans^H; the softmax-normalized aggregation weights in Eq.

more of objection 2

(10) are not orthonormal, so the claim that the optimal reconstruction is "strictly proportional to the adjoint operator" is false as stated and is also in tension with the appendix's own equation W_opt proportional to W_trans^dagger. The argument that the missing scaling is absorbed by normalization layers is a reasonable design rationale, but it should be presented as an inductive bias rather than as a proof, and the main-text sentence "This design is fundamentally rooted in the Best Approximation Theorem ... The proof is provided in Appendix A.2" should be qualified accordingly, as should the [Yes] answer to checklist item 3.

This review says the paper's mathematics is wrong. Pith did not check that against the paper's source.

The paper says, at Section 5.1 and Tables 1-3, 8

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

Section 5.1 states that all reported metrics are averaged over multiple independent runs, and checklist item 7 is answered [Yes] on statistical significance, but no standard deviation, confidence interval, or significance test appears anywhere in the paper. For the headline comparisons this is a minor loss because the margins are large, but several load-bearing comparisons are close: the ablation in …

more of objection 3

… Table 3 (DIAGNO MAE 0.1638 vs m-dim 0.1738), the ENSO phase table (Table 8, DIAGNO El Nino RMSE 0.0535 vs SFNO 0.0408), and the 10-day ocean MAE (0.1638 vs GSNO 0.1726). The claim of consistently achieving state-of-the-art performance needs at least mean +/- standard deviation over the independent runs, or a paired significance test, for these tables.

The paper says, at Appendix B.3 and Table 2

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The comparison against the foundation models is run under asymmetric training budgets, which matters for the claim of outperforming ClimaX and LangYa. Per Appendix B.3, ClimaX receives only 20 epochs of single-step fine-tuning, whereas all other models receive 100 epochs of one-step pretraining plus 100 epochs of two-step autoregressive fine-tuning, and LangYa is trained from scratch even though its architecture is designed for pretrained use.

more of objection 4

Please justify these budgets, for example by verifying that further two-step fine-tuning does not improve ClimaX on the 120-hour metrics, or restrict the comparison claim to the budgets actually evaluated.

6 minor objections

The paper says, at Section 5.2, Table 1

Pith says: minor objection 1 of 6

The source of the "35.2%" figure should be stated explicitly; it matches the mean of the four per-metric relative reductions versus GSNO (41.1%, 41.0%, 29.0%, 29.6%), not a reduction of a single averaged error, as the current phrasing suggests.

The paper says, at Section 4.2, Eq. (11)

Pith says: minor objection 2 of 6

Please specify whether the latent query q_k depends on the input state or is a static learned vector, and define what "refined by a self-attention layer" means; this directly affects how the attention matrices in Section 5.4 should be interpreted.

show 4 more

The paper says, at Section 4.2, Eqs. (13)-(14)

Pith says: minor objection 3 of 6

Since the m < 0 coefficients are omitted in Section 3.1 by conjugate symmetry, state explicitly how the reconstruction step produces the negative-order coefficients needed by the inverse transform.

The paper says, at Table 8

Pith says: minor objection 4 of 6

The "Total RMSE" column is numerically equal to the sum of the three phase-wise RMSE values (0.0579 + 0.0535 + 0.0140 = 0.1254), which is not a valid aggregation of RMSE values; please clarify how the total is computed and consider reporting a combined RMSE over all samples instead.

The paper says, at References

Pith says: minor objection 5 of 6

Reference [33] contains apparent placeholder tokens in the author list ("LiweiNo", "XiaoxuLin") and should be checked against the published version; reference [14] also contains a spacing artifact ("V o").

The paper says, at Section 5.5, Table 3

Pith says: minor objection 6 of 6

Please describe how the l-dim and m-dim ablation variants are instantiated, in particular which spectral axis is aggregated and how, so that the ablation is reproducible.

What would settle it

Independently reproduce the SSWE experiment (Table 1) using the provided code and default hyperparameters, computing the 10-hour MAE and RMSE for DIAGNO. If the reproduced values fall within 5% of the reported values (0.0790 and 0.0985 respectively), the central empirical claim is verified.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 13 Nov 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic partly agrees 9 s · 31,432 tokens read, 713 written
  3. Editor wrote the verdict above 11 s · 32,298 tokens read, 593 written
  4. Referee minor revision: 4 major objections, 6 minor 1 min 58 s · 31,746 tokens read, 18,844 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper is worth your time. The core idea—replacing the rotation-equivariance assumption in spherical neural operators with an explicit cross-modal interaction kernel, then factorizing that kernel along the diagonals of the spherical harmonic spectrum to capture zonal and meridional dynamics—is genuinely novel and physically motivated. The derivation from the general integral operator to the factorized form (Equations 7–9) is mathematically sound, and the diagonal factorization is a clever inductive bias that the authors motivate from geophysical principles and test empirically.

The experimental evaluation is the paper's strongest asset. They run three distinct benchmarks (simulated shallow water, real-world ocean reanalysis, real-world atmospheric reanalysis) against a wide set of baselines including standard neural operators, spherical operators, and domain-specific foundation models. The results consistently show DIAGNO outperforming alternatives, often by a clear margin. The additional analyses on climatological anomalies and physical consistency (energy spectra) go beyond standard metrics and demonstrate practical value. The ablation study confirms that the diagonal structure matters more than simpler l-dim or m-dim alternatives.

The soft spots are real but not fatal. The most significant is the absence of explicit error bars or confidence intervals in the main tables. The paper states that metrics are averaged over multiple independent runs, but without reporting variance, it's hard to assess whether the improvements are statistically significant. This is a reporting gap, not a flaw in the architecture. Second, the diagonal factorization is an inductive bias that could miss important cross-modal interactions that don't align with the diagonal structure. The authors acknowledge this implicitly by motivating it physically, but they don't explore failure cases or regimes where the assumption might break down. Third, the reliance on complete global grids (due to SHT) is a genuine limitation that the authors state clearly in Section 6—it's not a hidden flaw.

The stress-test note is correct: there is no load-bearing flaw. The paper's central claim is well-supported. The missing error bars are the main thing I'd want clarified before citing the results.

Who this is for: anyone working on neural operators for spherical domains, global weather/climate modeling, or spectral methods for PDEs. It deserves a serious referee. I would accept a review invitation.

Recommendation: Accept with minor revision. The authors should add error bars or confidence intervals to their main results tables. The paper is otherwise solid.

The strongest objection

After careful review, I do not identify a load-bearing concern that would undermine the central claim. The paper's core contribution is a new diagonal factorization of the spectral interaction kernel for spherical neural operators, motivated by geophysical intuition about zonal and meridional dynamics. The derivation from the general integral operator to the factorized form (Equations 7-9) is mathematically sound. The diagonal structure is not an arbitrary restriction but a deliberate inductive bias that the authors motivate physically and test empirically. The experimental evaluation is comprehensive across three distinct benchmarks (SSWE, sea surface, atmosphere) with multiple baselines, and the results consistently show DIAGNO outperforming alternatives. The paper acknowledges its main limitation (requirement for complete global grids due to SHT) in Section 6. The reader's weakest_assumption about the diagonal factorization being 'not derived from first principles' is accurate but does not constitute a flaw: the factorization is a design choice that is tested empirically, and the paper does not claim it is derived from first principles. The missing error bars noted by the reader are a reporting concern rather than a load-bearing flaw in the argument. The paper states in Section 5.1 that 'All reported metrics are averaged over multiple independent runs to ensure reliability,' which partially addresses this, though explicit error bars would strengthen the presentation. Overall, the central claim that DIAGNO achieves state-of-the-art performance by explicitly parameterizing cross-modal spectral interactions is well-supported by the evidence presented.

The reader's verdict is CONDITIONAL with high confidence, citing the diagonal factorization as the weakest assumption and the lack of error bars as the main concern. I agree that the diagonal factorization is the key inductive bias, but I do not see it as a load-bearing weakness: it is a motivated design choice that is empirically validated. The paper acknowledges that a dense kernel would allow arbitrary interactions but argues the diagonal structure is physically grounded. This is a reasonable architectural decision, not a flaw. The missing error bars are a legitimate concern for reproducibility, but the paper states metrics are averaged over multiple runs, and the code is provided. The reader's concern about 'state-of-the-art' claims being tempered due to potential baseline hyperparameter tuning is valid but applies to most empirical ML papers and does not uniquely undermine this work. The paper provides detailed hyperparameter configurations in Table 4 and training procedures in Appendix B.3, which is more transparent than many comparable papers. I find the central argument well-supported and the empirical evidence convincing. The limitation about global grids is honestly stated. Therefore, I recommend the verdict remain UNCHANGED (CONDITIONAL), as the reader's conditional acceptance based on clarifying error bars is appropriate. My agreement with the reader is partial: I share the view that error bars would strengthen the paper, but I do not consider the diagonal factorization assumption to be a significant risk given the empirical validation.

The referee's report

If the reported results hold, DIAGNO is a meaningful advance for spherical spectral operator learning: it replaces the modal isolation of equivariant spherical convolutions with a structured, input-dependent modal interaction kernel at complexity O(C*N^1.5), dominated by the unavoidable SHT. The derivations in Section 4.1 are clean, and the paper is honest about scope: Appendix A.1 explicitly limits the modal-isolation analysis to strictly equivariant linear spherical convolutions and credits GSNO with a non-equivariant Green's-function formulation, and Section 6 states the complete-global-grid limitation. The experiments are unusually comprehensive for this area, and the margins over the strongest spherical baseline are large, e.g., roughly 29%-41% per-metric error reduction on SSWE relative to GSNO. The paper also ships code and dataset-generation instructions, which strengthens reproducibility. The concerns I raise are about the physical-interpretation narrative, an overstated optimality proof in the appendix, and missing variance reporting; these are local and fixable and do not, in my assessment, undermine the empirical central claim.

Pith's reading of the paper (a machine's summary, not the authors')

This paper introduces DIAGNO, a neural operator for forecasting Earth's atmospheric and oceanic dynamics on a sphere. Existing spherical neural operators assume rotation equivariance, which treats the Earth as a homogeneous system and prevents interactions between different spectral modes. DIAGNO removes this assumption by directly parameterizing cross-modal interactions in the spherical harmonic domain. The authors introduce a diagonal factorization of the interaction kernel that respects the zonal and meridional structure of global dynamics, enabling the model to capture heterogeneous physical processes. Experiments across shallow water, sea surface, and atmospheric benchmarks show consistent state-of-the-art forecasting accuracy, with the model also demonstrating skill in capturing climate anomalies and maintaining physical energy spectra.

What carries the argument. The central machinery is the DIAGNO module, which operates in the spherical harmonic spectral domain. It factorizes the cross-modal interaction kernel W into three operations: an intra-diagonal transformation (W_trans) that aggregates modes sharing the same zonal pattern into a unified zonal mode, an inter-diagonal evolution operator (A_attn) that uses attention to redistribute energy among different zonal patterns, and an intra-diagonal conjugate reconstruction (W_rec) that maps the evolved zonal modes back to the spectral domain using the conjugate of the aggregation weights. This factorization is grounded in the observation that the diagonal structure of the spherical harmonic spectrum naturally characterizes zonal and meridional dynamics.

Core claim. The paper claims that explicitly parameterizing cross-modal interactions in the spherical spectral domain, rather than relying on the rotation equivariance assumption that isolates spectral modes, yields a neural operator that better captures the heterogeneous dynamics of the Earth system. The key discovery is that the diagonal structure of the spherical harmonic spectrum can be leveraged to factorize the interaction kernel into intra-diagonal (meridional) and inter-diagonal (zonal) operations, providing an efficient and physically grounded way to model global-scale dynamics without resorting to a dense, unconstrained kernel.

The premise it rests on. The diagonal factorization of the interaction kernel (Equation 9) assumes that the dominant physical interactions in Earth's dynamics can be captured by separating intra-diagonal (meridional) and inter-diagonal (zonal) operations, and that this fixed grouping is sufficient to model heterogeneous dynamics across all scales and variables.

If the paper is right (Pith's inferences, not the authors')

If correct, DIAGNO provides a new standard for spherical neural operators that can be applied to any global-scale geophysical forecasting task, including weather, ocean, and climate prediction, with demonstrably better accuracy than rotation-equivariant alternatives.

The explicit parameterization of cross-modal spectral interactions suggests that other neural operator architectures on spheres could benefit from relaxing the rotation equivariance assumption and adopting structured, physically motivated kernel factorizations.

The model's demonstrated ability to capture climate anomalies (ENSO phases) and maintain physical consistency (kinetic energy spectra) indicates that data-driven neural operators can serve as reliable tools for climate analysis and process studies, not just forecasting.

The diagonal factorization provides a computationally efficient way to model heterogeneous dynamics, with complexity scaling as O(C·N^1.5), dominated by the spherical harmonic transforms, making it feasible for operational high-resolution global models.

The diagonal factorization, while physically motivated, is a strong inductive bias. A natural extension would be to allow the model to learn the optimal factorization structure (e.g., learnable diagonal groupings or a learnable sparsity pattern) rather than fixing it to the l-m+1 = k grouping. This could reveal whether the true dynamics align exactly with this grouping or if cross-diagonal interac

The model's reliance on complete global grids due to the spherical harmonic transform limits its application to regional or limited-area domains. Combining the diagonal interaction paradigm with localized spectral representations (e.g., windowed spherical harmonics or wavelets) could extend its utility to regional forecasting while preserving the benefits of explicit cross-modal interaction.

The conjugate reconstruction design, justified by the Best Approximation Theorem, forces the network to treat the forward and backward projections as a nearly lossless geometric pair. This inductive bias could be tested by comparing against a version with learnable reconstruction parameters; if DIAGNO still outperforms, it suggests that enforcing this structural constraint is beneficial for genera

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Accurate forecasting of Earth's dynamic systems is vital in geophysics, where neural operators offer a powerful data-driven paradigm. Standard operators often treat the Earth as a flat Euclidean plane, introducing geometric artifacts. In contrast, Spherical Neural Operators (SNOs) resolve this mismatch by operating within the spherical spectral space, naturally adapting to spherical geometries. However, for computational efficiency, existing SNOs typically rely on the rotation equivariance assumption. This assumption simplifies the transitions of physical states into isolated spectral modal evolutions, thereby over-idealizing the Earth as a homogeneous system. To solve this problem, we no longer rely on the rotation equivariance assumption that isolates spectral modal evolution, but directly derive an explicit modal interaction kernel and propose Diagonal Spherical Neural Operators named DIAGNO. DIAGNO explicitly parameterizes cross-modal interactions through spectral modal decomposition, thereby preserving the inherent heterogeneity of Earth dynamics. Furthermore, leveraging the unique diagonal structure of the spherical spectrum, we design inter- and intra-diagonal interaction mechanisms to natively capture essential zonal and meridional dynamics, respecting intrinsic geophysical regularities. Extensive experiments across simulated fluid dynamics and real-world Earth system reanalysis (including atmosphere and ocean) datasets demonstrate DIAGNO's superior capability, consistently achieving state-of-the-art performance in multi-step forecasting. Beyond this, we validate DIAGNO's practical analytical value in professional geophysical domains by assessing its fidelity in capturing climatological anomalies and maintaining physical consistency.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-13

The authors' main claim, as the classifier picked it out 0.98

“Extensive experiments across simulated fluid dynamics and real-world Earth system reanalysis (including atmosphere and ocean) datasets demonstrate DIAGNO's superior capability, consistently achieving state-of-the-art performance in multi-step forecasting.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.09 · human participants 0.02 · controlled comparison 0.42 · randomized assignment 0.01 · states quantitative result 0.06 · compares with prior results 0.89 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Accurate forecasting of Earth's dynamic systems is vital in geophysics, where neural operators offer a powerful data-driven paradigm. background 1.00 0.04 0.03 asserted 1.00
Standard operators often treat the Earth as a flat Euclidean plane, introducing geometric artifacts. background 0.73 0.05 0.02 asserted 1.00
In contrast, Spherical Neural Operators (SNOs) resolve this mismatch by operating within the spherical spectral space, naturally adapting to spherical geometries. background 1.00 0.06 0.04 asserted 1.00
However, for computational efficiency, existing SNOs typically rely on the rotation equivariance assumption. background 0.85 0.05 0.02 asserted 0.98
This assumption simplifies the transitions of physical states into isolated spectral modal evolutions, thereby over-idealizing the Earth as a homogeneous system. gap 0.47 0.08 0.04 asserted 1.00
To solve this problem, we no longer rely on the rotation equivariance assumption that isolates spectral modal evolution, but directly derive an explicit modal interaction kernel and propose Diagonal Spherical Neural Operators named DIAGNO. contribution 0.89 0.96 0.23 asserted 0.99
DIAGNO explicitly parameterizes cross-modal interactions through spectral modal decomposition, thereby preserving the inherent heterogeneity of Earth dynamics. method 0.66 0.89 0.24 asserted 1.00
Furthermore, leveraging the unique diagonal structure of the spherical spectrum, we design inter- and intra-diagonal interaction mechanisms to natively capture essential zonal and meridional dynamics, respecting intrinsic geophysical regularities. method 0.85 0.90 0.07 asserted 1.00
Extensive experiments across simulated fluid dynamics and real-world Earth system reanalysis (including atmosphere and ocean) datasets demonstrate DIAGNO's superior capability, consistently achieving state-of-the-art performance in multi-step forecasting. result 1.00 0.96 0.54 asserted 1.00
Beyond this, we validate DIAGNO's practical analytical value in professional geophysical domains by assessing its fidelity in capturing climatological anomalies and maintaining physical consistency. result 0.73 0.94 0.13 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from simulation or computation 0.73

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-13: 1 request, 7,643 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (13)
Figure 1 of the paper
Figure 1. Comparison in spherical spectral space. Spherical Harmonics form a triangular (l, m) spectrum. Un￾like previous SNOs using element-wise transforms, DIAGNO enables intra- and inter-diagonal interactions. To address the limitation, we propose the Diagonal Spher￾ical Neural Operator named DIAGNO1 , which shifts the modeling paradigm from isolated modal evolution to ex￾plicit cross-modal interaction. Rather than relying on ide￾alized system modeling, DIAGNO utilizes direct spectral modal decomposition to derive a parameterized interaction kernel, thereby effectively capturing the…
Figure 2 of the paper
Figure 2. 10-hour SSWE velocity field (U) forecasting from a North Pole perspective. (a) Ground truth. Standard modules (b-c) exhibit radial artifacts at the pole. SNOs (d-f) eliminate artifacts, though (d-e) suffer from energy dissipation. (f) DIAGNO maintains the highest physical fidelity. graph frameworks to resolve multi-scale features. However, when applied to Earth system modeling, these architectures fail to respect the inherent spherical geometry. Operating within Euclidean planes, they introduce spatial aliasing that manifests as radial artifacts near the poles. In contrast,…
Figure 3 of the paper
Figure 3. DIAGNO module architecture. Complex spectral coefficients are represented as triangular matrices and evolved through sequential intra-diagonal transformation, inter-diagonal evolution, and intra-diagonal conjugate reconstruction. Spheres illustrate spatial patterns of spectral modes. where uˆl ′ ,m′ are spectral modal coefficients of u(x ′ ). Through this transformation, we define the kernel of modal interaction W(l,m),(l ′ ,m′) , which encapsulates interaction between input spectral mode (l ′ , m′ ) and target spectral mode (l, m). Therefore, the continuous spatial operator in…
Figure 4 of the paper
Figure 4. The ACC of ocean surface prediction (higher is better). DIAGNO exhibits superior stability across all lead times compared to spherical operators and the domain-specific LangYa model. 7
Figure 5 of the paper
Figure 5. SSTA RMSE comparison. Left axis: phase-wise SSTA RMSE (bars). Right axis: total SSTA RMSE (line) on a logarithmic scale for visual compactness. See
Figure 6 of the paper
Figure 6. Physical consistency analysis. X￾axis: Spatial EKE RMSE. Y-axis: Log KES RMSE on a logarithmic scale. ClimaX shows competitive predictive performance second to DIAGNO, yet falls considerably short in energy consistency. To verify practical analytical value, we extend our evaluation beyond general metrics to focus on two advanced dimensions: sensitivity to climatological anomalies and physical consistency in energetics. Climatological Anomaly. We evaluate climate anomaly capturing via 7-day Sea Surface Temper￾ature Anomaly (SSTA) forecasts in the Niño 3.4 region [61] (
Figure 7 of the paper
Figure 7. Scale-specific dominant mode reconstructions. (a) Actual SST field. (b-f) iSHT representations reveal hierarchical extraction: (b) lowest-resolution interhemispheric asymmetry; (c-d) intermediate-resolution climate belts; (e-f) high-resolution fine asymmetric details. 5.4 Interpreting Dominant Dynamics: A Scale-Resolution Correspondence (RQ3) To verify how DIAGNO captures dominant dynamics across resolutions, we analyze its cross-attention matrices generated during Sea Surface Temperature (SST) forecasting, which quantify the coupling between aggregated zonal modes and latent…
Figure 8 of the paper
Figure 8. Hyperparameter analysis. 10-hour prediction RMSE on SSWE. Our configuration (5 multi-resolution layers, 128 dim) is optimal. Conclusions. We propose DIAGNO, which abandons the over-idealized rotation equivariance of previous spherical neural operators to explic￾itly model cross-modal interactions. By leverag￾ing the unique diagonal structure of the Spherical Harmonic spectrum, DIAGNO natively captures global-scale zonal and meridional dynamics, pre￾serving the intrinsic heterogeneity of Earth sys￾tems. Extensive evaluations across atmospheric and oceanic scenarios demonstrate that…
Figure 9 of the paper
Figure 9. Comparison of forecasting results for SSWE from a north pole perspective.
Figure 10 of the paper
Figure 10. Day-to-day forecasting performance of different models on global ocean surface variables
Figure 11 of the paper
Figure 11. Spatial visualizations of 10-day absolute prediction errors for four key ocean surface
Figure 12 of the paper
Figure 12. Autoregressive forecasting performance of different models on global atmospheric
Figure 13 of the paper
Figure 13. Spatial visualizations of 20-step (5-day) absolute prediction errors for three key atmo

Figures are the authors', reproduced from the paper.

References (67)
  1. [1] Salva Rühling Cachay, Brian Henn, Oliver Watt-Meyer, Christopher S. Bretherton, and Rose Yu. Prob- ablistic emulation of a global climate model with spherical dyffusion. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in N... on Pith
  2. [2] Qiusheng Huang, Yuan Niu, Xiaohui Zhong, Anboyu Guo, Lei Chen, Dianjun Zhang, Xuefeng Zhang, and Hao Li. Fuxi-ocean: A global ocean forecasting system with sub-daily resolution, 2025. URL https://arxiv.org/abs/2506.03210 on PitharXiv
  3. [3] Yuze Sun, Wentao Luo, Yanfei Xiang, Jiancheng Pan, Jiahao Li, Quan Zhang, and Xiaomeng Huang. Deploying atmospheric and oceanic ai models on chinese hardware and framework: Migration strategies, performance optimization and analysis. InProceedings of the AAAI Conference on Art... on Pith
  4. [4] Z. Wang, B. Lu, Y . Xin, T. Ito, L. Zhou, L. Cheng, Y . Li, X. Wang, and M. Jin. Jingwei-nutrients: a global spatiotemporal reconstruction of ocean nutrients (1965–2023) using multi-task deep learning. Earth System Science Data, 18(7):5069–5092, 2026. doi: 10.5194/essd-18-5069... on Pithdoi:10.5194/essd-18-5069-2026
  5. [5] Bin Lu, Jingjing Shen, Ze Zhao, Jianping Zhou, Haonan Qi, Luyu Han, Xiaoying Gan, Meng Jin, Lei Zhou, Luoyi Fu, Xinbing Wang, and Chenghu Zhou. Oceanverse: Evaluable 4d ocean element reconstruction dataset under realistic sparsity. InProceedings of the 32nd ACM SIGKDD Conferen... on Pithdoi:10.1145/3770855.3817534
  6. [6] Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567):47–55, 2015 on Pith
  7. [7] Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew M. Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In9th International Conference on Learning Representations, ICLR 20... on Pith
  8. [8] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature machine intelligence, 3(3):218–229, 2021 on Pith
  9. [9] Andrew Staniforth and John Thuburn. Horizontal grids for global weather and climate prediction models: a review.Quarterly Journal of the Royal Meteorological Society, 138(662):1–26, 2012 on Pith
  10. [10] Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical Fourier neural operators: Learning stable dynamics on the sphere. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Saba... on Pith
  11. [11] Hao Tang, Hao Chen, and Chao Li. Generalized spherical neural operators: Green’s function for- mulation. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=XkGjzSDTnm on Pith
  12. [12] Yifan Hu, Fukang Yin, Weimin Zhang, Kaijun Ren, Junqiang Song, Kefeng Deng, and Di Zhang. Spherical multigrid neural operator for improving autoregressive global weather forecasting.Scientific Reports, 15 (1):11522, 2025. 10 on Pith
  13. [13] Taco S Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical cnns.arXiv preprint arXiv:1801.10130, 2018 on PithPith's reviewarXiv
  14. [14] Trond H Torsvik, Rob van der V oo, Pavel V Doubrovine, Kevin Burke, Bernhard Steinberger, Lewis D Ashwal, Reidar G Trønnes, Susan J Webb, and Abigail L Bull. Deep mantle structure as a reference frame for movements in and on the earth.Proceedings of the National Academy of Sci... on Pith
  15. [15] Adam M Dziewonski, Vedran Lekic, and Barbara A Romanowicz. Mantle anchor structure: an argument for bottom up tectonics.Earth and Planetary Science Letters, 299(1-2):69–79, 2010 on Pith
  16. [16] Brian J Hoskins and David J Karoly. The steady linear response of a spherical atmosphere to thermal and orographic forcing.Journal of Atmospheric Sciences, 38(6):1179–1196, 1981 on Pith
  17. [17] Tapio Schneider. The general circulation of the atmosphere.Annu. Rev. Earth Planet. Sci., 34(1):655–688, 2006 on Pith
  18. [18] Haonan Qi, Bin Lu, Yimian Hu, Ze Zhao, Meng Jin, Lixin Qu, Shuai Li, Shizhen Zhao, Lei Zhou, Xiaoying Gan, Xinbing Wang, and Chenghu Zhou. From surface to depth: Diagnosing ocean vertical velocity with 3d frequency operator. InProceedings of the 32nd ACM SIGKDD Conference on K... on Pithdoi:10.1145/3770855.3819028
  19. [19] John Guibas, Morteza Mardani, Zongyi Li, Andrew Tao, Anima Anandkumar, and Bryan Catanzaro. Adaptive fourier neural operators: Efficient token mixers for transformers.CoRR, abs/2111.13587, 2021. URLhttps://arxiv.org/abs/2111.13587 on PithPith's reviewarXiv
  20. [20] Gege Wen, Zongyi Li, Kamyar Azizzadenesheli, Anima Anandkumar, and Sally M Benson. U-fno—an enhanced fourier neural operator-based deep-learning model for multiphase flow.Advances in Water Resources, 163:104180, 2022 on Pith
show 47 more
  1. [21] Alasdair Tran, Alexander Patrick Mathews, Lexing Xie, and Cheng Soon Ong. Factorized fourier neural operators. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Ki- gali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. URL https://openreview.net/foru... on Pith
  2. [22] Md Ashiqur Rahman, Zachary E. Ross, and Kamyar Azizzadenesheli. U-NO: u-shaped neural operators. Trans. Mach. Learn. Res., 2023, 2023. URLhttps://openreview.net/forum?id=j3oQF9coJd on Pith
  3. [23] Jean Kossaifi, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, and Anima Anandkumar. Multi-grid tensorized fourier neural operator for high- resolution pdes.Trans. Mach. Learn. Res., 2024, 2024. URL https://openreview.net/forum?id=AWiDlO63bH on Pith
  4. [24] Shuhao Cao. Choose a transformer: Fourier or galerkin. In Marc’Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, editors,Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Syste... on Pith
  5. [25] Xinliang Liu, Bo Xu, and Lei Zhang. Ht-net: Hierarchical transformer based operator learning model for multiscale pdes.CoRR, abs/2210.10890, 2022. doi: 10.48550/ARXIV .2210.10890. URL https: //doi.org/10.48550/arXiv.2210.10890 on PithPith's reviewarXivdoi:10.48550/arxiv.2210.10890
  6. [26] Zijie Li, Kazem Meidani, and Amir Barati Farimani. Transformer for partial differential equations’ operator learning.Trans. Mach. Learn. Res., 2023, 2023. URL https://openreview.net/forum?id= EPPqt3uERT on Pith
  7. [27] Zijie Li, Dule Shu, and Amir Barati Farimani. Scalable transformer for pde surrogate modeling.Advances in Neural Information Processing Systems, 36:28010–28039, 2023 on Pith
  8. [28] Zhongkai Hao, Zhengyi Wang, Hang Su, Chengyang Ying, Yinpeng Dong, Songming Liu, Ze Cheng, Jian Song, and Jun Zhu. GNOT: A general neural operator transformer for operator learning. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonatha... on Pith
  9. [29] Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for pdes on general geometries. In Ruslan Salakhutdinov, Zico Kolter, Katherine A. Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors,F... on Pith
  10. [30] Zipeng Xiao, Zhongkai Hao, Bokai Lin, Zhijie Deng, and Hang Su. Improved operator learning by orthogonal attention. In Ruslan Salakhutdinov, Zico Kolter, Katherine A. Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors,Forty-first Internationa... on Pith
  11. [31] Hang Zhou, Yuezhou Ma, Haixu Wu, Haowen Wang, and Mingsheng Long. Unisolver: PDE-conditional transformers towards universal neural PDE solvers. InForty-second International Conference on Machine Learning, 2025. URLhttps://openreview.net/forum?id=r1ryQoI9iZ on Pith
  12. [32] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. InThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. URL https://openreview.net/forum?id... on Pith
  13. [33] Jianwei Zheng, LiweiNo, Ni Xu, Junwei Zhu, XiaoxuLin, and Xiaoqin Zhang. Alias-free mamba neural operator. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=gUEBXGV8JM on Pith
  14. [34] Nodens F. Koren and Samuel Lanthaler. Merging memory and space: A state space neural operator.Trans. Mach. Learn. Res., 2026, 2026. URLhttps://openreview.net/forum?id=SwLxxz0x58 on Pith
  15. [35] Yanan Guo, Junqiang Song, Xiaoqun Cao, Chuanfeng Zhao, Fukang Yin, and Hongze Leng. PGMNO: A physics-guided mamba neural operator framework for partial differential equations.Neural Networks, 201: 108845, 2026. doi: 10.1016/J.NEUNET.2026.108845. URL https://doi.org/10.1016/j.n... on Pithdoi:10.1016/j.neunet.2026.108845
  16. [36] Zeyuan Song and Zheyu Jiang. Adaptive mamba neural operators. InThe Fourteenth International Conference on Learning Representations, 2026 on Pith
  17. [37] Zongyi Li, Nikola B. Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Andrew M. Stuart, Kaushik Bhattacharya, and Anima Anandkumar. Multipole graph neural operator for parametric partial differential equations. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria- Flor... on Pith
  18. [38] Zongyi Li, Nikola B. Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew M. Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial differential equations. CoRR, abs/2003.03485, 2020. URLhttps://arxiv.org/abs/2003.03485 on PithPith's reviewarXiv
  19. [39] Zhihao Li, Haoze Song, Di Xiao, Zhilu Lai, and Wei Wang. Harnessing scale and physics: A multi-graph neural operator framework for pdes on arbitrary geometries. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V .1, KDD ’25, page 729–740, ... on Pith
  20. [40] Wen Ye, Wei Yang, Defu Cao, Yizhou Zhang, Lumingyuan Tang, Jie Cai, and Yan Liu. TS-reasoner: Domain-oriented time series inference agents for reasoning and automated analysis.Transactions on Machine Learning Research, 2026. ISSN 2835-8856. URL https://openreview.net/forum?id=... on Pith
  21. [41] Defu Cao, Wen Ye, Yizhou Zhang, Sam Griesemer, and Yan Liu. PINFDit: Energy-based physics-informed diffusion transformers for general-purpose time series tasks. InThe Fourteenth International Conference on Learning Representations, 2026. URLhttps://openreview.net/forum?id=EphTlUJ4XN on Pith
  22. [42] Kendall Atkinson and Weimin Han.Spherical harmonics and approximations on the unit sphere: an introduction, volume 2044. Springer Science & Business Media, 2012. 12 on Pith
  23. [43] Milton Abramowitz and Irene A Stegun.Handbook of mathematical functions with formulas, graphs, and mathematical tables, volume 55. US Government printing office, 1948 on Pith
  24. [44] Harry Nyquist. Certain topics in telegraph transmission theory.Transactions of the American Institute of Electrical Engineers, 47(2):617–644, 1928 on Pith
  25. [45] CE Shannon. Communication in the presence of noise.Proceedings of the IRE, 37(1):10–21, 1949 on Pith
  26. [46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells III, and Alejandro F. Frangi, editors,Medical Image Computing and Computer-Assisted Intervention - MICCAI 2... on Pithdoi:10.1007/978-3-319-24574-4
  27. [47] Susanne C Brenner and L Ridgway Scott.The mathematical theory of finite element methods, volume 15. Springer, 2008 on Pith
  28. [48] Isaac M Held and Arthur Y Hou. Nonlinear axially symmetric circulations in a nearly inviscid atmosphere. Journal of the Atmospheric Sciences, 37(3):515–533, 1980 on Pith
  29. [49] Kevin E Trenberth and Julie M Caron. Estimates of meridional atmosphere and ocean heat transports. Journal of Climate, 14(16):3433–3443, 2001 on Pith
  30. [50] Alexandre Ganachaud and Carl Wunsch. Improved estimates of global ocean circulation, heat transport and mixing from hydrographic data.Nature, 408(6811):453–457, 2000 on Pith
  31. [51] Erwin Kreyszig.Introductory functional analysis with applications. John Wiley & Sons, 1991 on Pith
  32. [52] Haim Brezis and Haim Brézis.Functional analysis, Sobolev spaces and partial differential equations, volume 2. Springer, 2011 on Pith
  33. [53] Ramachandran D Nair, Stephen J Thomas, and Richard D Loft. A discontinuous galerkin global shallow water model.Monthly weather review, 133(4):876–888, 2005 on Pith
  34. [54] Boris Bonev, Jan S Hesthaven, Francis X Giraldo, and Michal A Kopera. Discontinuous galerkin scheme for the spherical shallow water equations with applications to tsunami modeling and prediction.Journal of Computational Physics, 362:425–448, 2018 on Pith
  35. [55] Lellouche Jean-Michel, Greiner Eric, Bourdallé-Badie Romain, Garric Gilles, Melet Angélique, Drévillon Marie, Bricaud Clément, Hamon Mathieu, Le Galloudec Olivier, Regnier Charly, et al. The copernicus global 1/12 oceanic and sea ice glorys12 reanalysis.Frontiers in Earth Scie... on Pith
  36. [56] Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquín Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis.Quarterly journal of the royal meteorological society, 146(730):1999–2049, 2020 on Pith
  37. [57] Haixu Wu, Tengge Hu, Huakun Luo, Jianmin Wang, and Mingsheng Long. Solving high-dimensional pdes with latent spectral models. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors,International Conference on Machine ... on Pith
  38. [58] Nan Yang, Chong Wang, Meihua Zhao, Zimeng Zhao, Huiling Zheng, Bin Zhang, Jianing Wang, and Xiaofeng Li. Langya: Revolutionizing cross-spatiotemporal ocean forecasting.CoRR, abs/2412.18097, on PithPith's reviewarXiv
  39. [59] Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K. Gupta, and Aditya Grover. Climax: A foundation model for weather and climate. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors,International Conferenc... on Pith
  40. [60] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In2021 IEEE/CVF International Con- ference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2... on Pithdoi:10.1109/iccv48922.2021.00986
  41. [61] Kevin E Trenberth. The definition of el nino.Bulletin of the American Meteorological Society, 78(12): 2771–2778, 1997 on Pith
  42. [62] Boyin Huang, Peter W Thorne, Viva F Banzon, Tim Boyer, Gennady Chepurin, Jay H Lawrimore, Matthew J Menne, Thomas M Smith, Russell S V ose, and Huai-Min Zhang. Extended reconstructed sea surface temperature, version 5 (ersstv5): upgrades, validations, and intercomparisons.Jour... on Pith
  43. [63] KS Gage. A climatology of atmospheric wavenumber spectra of wind and temperature observed by commercial aircraft.J. Atmos. Sci, 42:950–960, 1985 on Pith
  44. [64] William C Skamarock. Evaluating mesoscale nwp models using kinetic energy spectra.Monthly weather review, 132(12):3019–3032, 2004 on Pith
  45. [65] Brian J Hoskins and Kevin I Hodges. New perspectives on the northern hemisphere winter storm tracks. Journal of the Atmospheric Sciences, 59(6):1041–1061, 2002. 14 A Theoretical Supplements A.1 The Mathematical Origin of Isolated Modes in Spherical Neural Operators To understa... on Pith
  46. [2024] doi: 10.48550/ARXIV .2412.18097. URLhttps://doi.org/10.48550/arXiv.2412.18097 on PithPith's reviewarXivdoi:10.48550/arxiv.2412.18097
  47. [2025] Association for Computing Machinery. ISBN 9798400712456. doi: 10.1145/3690624.3709173. URLhttps://doi.org/10.1145/3690624.3709173 on Pithdoi:10.1145/3690624.3709173

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment