DIAGNO: Diagonal Spherical Neural Operators for Heterogeneous Earth Dynamics Modeling
Authors: Herui Li, Bin Lu, Haonan Qi, Lei Zhou, Luoyi Fu, Xinbing Wang, Meng Jin
Organizations: School of Information Science and Electronic Engineering, Shanghai Jiao Tong University · School of Oceanography, Shanghai Jiao Tong University · School of Computer Science, Shanghai Jiao Tong University · School of Artificial Intelligence, Shanghai Jiao Tong University
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.
Figures & tables
Figure 1: Comparison in spherical spectral space. Spherical Harmonics form a triangular (l,m) spectrum. Unlike previous SNOs using element-wise transforms, DiagNO enables intra- and inter-diagonal interactions.
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.
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.
MODEL
#MEM
5 hours
10 hours
MAE ↓
RMSE ↓
MAE ↓
RMSE ↓
(GB)
( ×103 )
( ×103 )
( ×103 )
( ×103 )
Swin [ 60 ]
2.49
0.2285
0.2878
0.2545
0.3212
FNO [ 7 ]
2.50
0.2702
0.3390
0.2653
0.3326
LSM [ 57 ]
3.92
0.1038
0.1583
0.1905
0.2502
Transolver [ 29 ]
5.61
0.2897
0.3636
0.2871
0.3591
Table 1: Performance comparison for SSWE. Spherical neural operators generally outperform standard baselines. DiagNO achieves the best performance, reducing averaged error by 35.2% while maintaining competitive memory efficiency.
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.
MODEL
72 Hours (12 steps)
120 Hours (20 steps)
MAE ↓
RMSE ↓
ACC(%) ↑
MAE ↓
RMSE ↓
ACC(%) ↑
( ×103 )
( ×103 )
Z
U
V
Avg.
( ×103 )
( ×103 )
Z
U
V
Avg.
Swin [ 60 ]
0.2029
0.3014
68.97
59.98
55.36
61.44
0.3147
0.4627
38.53
33.21
27.17
32.97
FNO [ 7 ]
0.2974
0.4145
36.47
37.04
33.40
35.63
0.4710
0.6404
12.45
15.32
12.92
13.57
LSM [ 57 ]
0.1697
0.2503
73.64
62.36
58.28
64.76
0.2448
0.3595
46.28
37.86
31.99
38.71
Transolver [ 29 ]
0.2956
0.4165
36.60
27.89
23.95
29.48
0.3918
0.5344
17.53
12.05
11.43
13.67
Table 2: Performance comparison for ERA5 atmospheric forecasting. DiagNO consistently outperforms all baselines across all lead times, with particularly strong improvements in Z forecasting.
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 Table 8 for exact values.
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.
Table 4: Hyper-parameter configurations of different baseline models.
Module / Process
Computational Complexity
Spherical Harmonic Transform (SHT)
O(C⋅N1.5)
Intra-Diagonal Transformation
O(C⋅N)
Inter-Diagonal Evolution
O(C⋅N)
Intra-Diagonal Conjugate Reconstruction
O(C⋅N)
Inverse SHT (iSHT)
O(C⋅N1.5)
Overall DiagNO Module
O(C⋅N1.5)
Appendix
Table 5: Computational complexity of the DiagNO module. The complexity is measured with respect to the total number of spatial grid nodes N and the hidden channel dimension C .
MODEL
#MEM (GB)
SSWE
Sea Surface Forecasting
Atmospheric Forecasting
Swin
2.49
4.50
2.41
FNO
2.50
3.83
2.40
LSM
3.92
6.78
3.84
Transolver
5.61
12.28
5.54
MambaNO
12.14
26.59
12.06
Appendix
Table 6: Peak memory consumption (#MEM) of different models across three benchmark datasets.
Figure 9: Comparison of forecasting results for SSWE from a north pole perspective.
MODEL
MAE ↓
RMSE ↓
ACC(%) ↑
SSH
SST
U
V
Average
3 Days
Swin
0.2440
0.5262
52.12
26.20
45.15
37.50
40.24
FNO
0.2359
0.3173
80.09
56.90
80.64
76.81
73.61
LSM
0.1335
0.1758
89.98
83.78
90.15
88.85
88.19
Transolver
0.0972
0.1366
94.62
92.05
90.49
89.04
91.55
Appendix
Table 7: The overall performance comparison for ocean surface prediction.
Figure 10: Day-to-day forecasting performance of different models on global ocean surface variables over a 10-day lead time.
Figure 11: Spatial visualizations of 10-day absolute prediction errors for four key ocean surface variables. The corresponding overall Root Mean Square Error (RMSE) for each evaluated model is explicitly annotated beneath its respective spatial error map.
Figure 12: Autoregressive forecasting performance of different models on global atmospheric variables over a 5-day lead time. The x-axis represents the forecast horizon spanning 20 consecutive 6-hour steps.
Figure 13: Spatial visualizations of 20-step (5-day) absolute prediction errors for three key atmospheric variables. The corresponding overall Root Mean Square Error (RMSE) for each evaluated model is explicitly annotated beneath its respective spatial error map. Note that the annotated RMSE values for Z , U , and V are scaled by a factor of 103 , 10 , and 10 , respectively, for visual clarity.
MODEL
Total
RMSE
Neutral
El Niño
La Niña
Swin
0.8713
0.2618
0.5053
0.1042
FNO
1.4281
0.4227
0.7389
0.2665
LSM
1.8451
0.4924
0.6296
0.7231
Transolver
0.5427
0.1675
0.1376
0.2376
MambaNO
2.8335
0.8485
1.0422
0.9428
Appendix
Table 8: RMSE evaluation of different models across ENSO phases.
Meteorological downscaling is crucial for high-resolution regional climate forecasting and disaster early warning. While neural operators have emerged as a promising paradigm for modeling complex spatiotemporal mappings, existing frameworks often struggle with spherical manifold geometric distortions, inherent atmospheric multi-scale coupling mismatches, and lack of explicit atmospheric laws. We propose the Spherical Physics-informed Neural Operator, which utilizes a Spherical Laplacian Decomposition to partition atmospheric fields into hierarchical frequency components, maintaining exact point-wise correspondence across scales. To evaluate these representations at arbitrary locations, we introduce a localized spherical integral operator that approximates continuous kernel transforms via geometry-aware attention. Dynamical consistency is further enforced by embedding differentiable constraints into the learning process. Extensive experiments demonstrate that our framework attains superior accuracy and zero-shot generalization across various meteorological variables and unseen queries, representing a robust and interpretable solution for global-to-regional meteorological downscaling.
Yiqiang Ye, Yichi Wang, Haodi Zhang +5
College of Computer and Software Engineering, Shenzhen University, Shenzhen, China · School of Computer Science, Wuhan University, Wuhan, China
Spherical-harmonic neural operators provide an efficient representation for global forecasting, but their discretized pointwise nonlinearities can break rotational equivariance through aliasing. Clebsch-Gordan (CG) tensor products provide exactly equivariant interactions between spherical-harmonic representations, but their computational 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 practical, 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 forecast skill relative to SFNO on WeatherBench. Matched ablations show that CG coupling accelerates convergence, while SSHCG keeps the coupling to about 10 percent of a training step at the operating bandwidth, making exact equivariant coupling practical at forecasting scale.
Many dynamical processes unfold on the sphere but the default scientific machine learning architectures are Euclidean. Applying these architectures on a regular lat--lon grid causes problems: Cartesian convolutions become distorted at high latitude; 2D FFTs in Fourier neural operators incorrectly assume double periodicity; Cartesian positional encodings in ViTs distort spherical geodesic distances. Recent work moves towards natively spherical primitives, including spherical convolutions (e.g., DeepSphere or DISCO), Spherical Fourier Neural Operators (SFNOs), and geodesic attention. Here we propose Dandelion, a spherical version of Flower---a warp-based neural PDE solver. A Dandelion layer computes a tangent-plane displacement and then transports features along great circles. We obtain a U-Net-like structure by implementing hierarchical pooling entirely in the spherical-harmonic domain. There are thus no convolutions: spatial mixing is achieved only through spherical coordinate changes---or warps. To compare Dandelion with existing spherical architectures, we release an evolving benchmark suite of challenging, natively spherical PDE datasets including a modified Galewsky jet, anomalous chained turbulence, Cahn--Hilliard decomposition, spherical Riemann shocks, Held--Suarez dry atmospheric transport and global ocean dynamics. This new benchmark fills the gap between existing spherical datasets which are either too small and stylized, or much too large (ERA5) for model iteration. Dandelion achieves the best performance on all but one benchmark setting, where the best model is Flower2D, also based on warps. The spherical primitives result in the most physically consistent forecasts.