ClimateLLM: Efficient Weather Forecasting via Frequency-Aware Large Language Models

ICLR 2026 Conference Submission14599 Authors

Keywords: Spatiotemporal Modeling, Mixture-of-Experts, Weather Forecasting

Abstract: Recent progress in deep learning has advanced global weather forecasting, with larger and higher-resolution models steadily improving skill. In parallel, spectral methods provide an efficient basis for global dynamics. Yet most spectral approaches treat the complex spectrum as generic features, conflating the distinct physics encoded in amplitude (energy evolution) and phase (spatial propagation). **We propose ClimateLLM, a physics-aligned, frequency-domain forecasting framework powered by SAED-Former.** At its core, **SAED-Former** explicitly separates these two processes via a *dual-state representation*, computes interactions through a *phase-centric propagation kernel*, and injects wave-number–aware priors using *scale-conditional projection*. This physics-aligned design yields compact, robust frequency-domain representations. On standard reanalysis benchmarks, ClimateLLM matches or exceeds state-of-the-art accuracy across short- and medium-range horizons while training on a single GPU within hours. Moreover, the model supports *cross-variable transference*: networks trained on data-rich variables produce robust zero-shot forecasts for data-scarce variables. By elevating spectral structure to first-class status, ClimateLLM improves forecast quality, efficiency, and generalization.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 14599