DIAGNO: Diagonal Spherical Neural Operators for Heterogeneous Earth Dynamics Modeling

Herui Li, Bin Lu, Haonan Qi, Lei Zhou, Luoyi Fu, Xinbing Wang, Meng JinarXiv 2026 · arXiv

TL;DR

DIAGNO introduces a novel class of Spherical Neural Operators that abandons the restrictive "homogeneous Earth" assumption common in global forecasting models. By explicitly modeling cross-modal interactions through a unique diagonal factorization of the spherical harmonic spectrum, it captures the complex, location-dependent dynamics of our planet with state-of-the-art accuracy and physical fidelity.

What Problem Does This Paper Address?

Modeling global Earth systems (like the atmosphere and oceans) presents a unique topological challenge. Standard neural operators often treat the Earth as a flat Euclidean plane, which leads to "radial artifacts"—distortions near the poles where the coordinate system breaks down.

While existing Spherical Neural Operators (SNOs) solve this by working in the spherical spectral domain, they typically rely on a mathematical shortcut called rotation equivariance. This assumption implies that physical laws are the same regardless of your absolute location on the sphere. While computationally efficient, it effectively treats the Earth as a homogeneous, featureless ball. In reality, Earth is highly heterogeneous: continents, varying sea-surface temperatures, and the latitude-dependent Coriolis force mean that dynamics are tied to specific geographical locations. Current SNOs struggle to capture these localized, non-uniform interactions.

Methods

The authors propose DIAGNO (Diagonal Spherical Neural Operator), which shifts the paradigm from isolated modal evolution to explicit cross-modal interaction.

The core innovation lies in how they handle the Spherical Harmonic spectrum. In this spectral space, coefficients form a triangular $(l, m)$ structure. The authors observed that this structure isn't just a mathematical convenience; it aligns with physical regularities:

  • Intra-diagonal interactions: Modes along the same diagonal share zonal (east-west) patterns but differ in meridional (north-south) oscillations.
  • Inter-diagonal interactions: Moving across diagonals captures transitions between different zonal structures.
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.

The DIAGNO module processes data in three steps:

  1. Intra-Diagonal Transformation: Aggregates spectral modes sharing the same zonal pattern to capture meridional dynamics.
  2. Inter-Diagonal Evolution: Uses a physically-grounded attention mechanism to model energy transfer between different zonal patterns.
  3. Intra-Diagonal Conjugate Reconstruction: Maps the evolved states back to the full spectral domain using a mathematically optimized "conjugate" weight strategy derived from the Best Approximation Theorem.

Datasets

The model was evaluated on three distinct benchmarks:

  • Spherical Shallow Water Equations (SSWE): A simulated dataset for large-scale fluid dynamics on a rotating sphere.
  • GLORYS12: A real-world ocean reanalysis dataset tracking Sea Surface Height (SSH), Temperature (SST), and velocity (U, V).
  • ERA5: The "gold standard" atmospheric reanalysis dataset from ECMWF, focusing on Geopotential (Z) and wind fields (U, V) at the 500 hPa level.

Key Insights and Ideas

  • Breaking Equivariance is Key: By moving away from strict rotation equivariance, DIAGNO can "learn" that the physics near the equator is different from the physics near the poles or over land vs. ocean.
  • Physical-Spectral Mapping: The most elegant "aha" moment is the realization that the diagonal structure of spherical harmonics naturally encodes the orthogonal directions (zonal and meridional) of geophysical flow. This allows the model to respect geophysical regularities without hard-coding them.
  • Efficiency through Factorization: Modeling every possible interaction between spectral modes would be computationally explosive ($O(N^2)$). By factorizing the kernel into diagonal operations, DIAGNO maintains a complexity of $O(C \cdot N^{1.5})$, making it as efficient as standard SNOs while being far more expressive.

Key Results

DIAGNO consistently achieved state-of-the-art (SOTA) results across all benchmarks:

  • SSWE: Reduced averaged forecasting error by 35.2% compared to previous best models.
  • Ocean Forecasting: Maintained significantly higher stability in Anomaly Correlation Coefficient (ACC) over 10-day lead times compared to domain-specific models like LangYa.
  • Atmospheric Forecasting: Showed particularly strong improvements in Geopotential (Z) forecasting, a critical metric for capturing large-scale weather circulations.
  • Physical Fidelity: Unlike other models that suffer from energy dissipation (blurring), DIAGNO maintains sharp gradients and respects the "energy cascade" (how energy moves between large and small scales in the atmosphere).
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.

Limitations and Future Work

The primary limitation is that DIAGNO relies on the Spherical Harmonic Transform (SHT), which mathematically requires a complete global grid. This makes it difficult to apply the model to localized regional domains (e.g., just the North Atlantic).

The authors suggest that future work will focus on developing localized representations to independently model complex polar dynamics and extending the diagonal interaction paradigm to regional modeling. Additionally, while the model is highly performant, it remains a data-driven tool; integrating it more deeply with traditional physics-constrained numerical solvers could further enhance its reliability for extreme climate events.

Summary updated

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.
BibTeX
@misc{li2026diagno,
  author = {Herui Li and Bin Lu and Haonan Qi and Lei Zhou and Luoyi Fu and Xinbing Wang and Meng Jin},
  title = {{DIAGNO: Diagonal Spherical Neural Operators for Heterogeneous Earth Dynamics Modeling}},
  year = {2026},
  eprint = {2611.16356},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.16356},
}
Figure