Phys-Diff: A Physics-Inspired Latent Diffusion Model for Tropical Cyclone Forecasting

TL;DR

Phys-Diff is a novel physics-inspired latent diffusion model that improves tropical cyclone forecasting by embedding physical relationships among attributes using multimodal data and attention mechanisms.

Contribution

It introduces a disentangled latent diffusion framework with physics-inspired biases and multimodal data integration for more physically consistent cyclone predictions.

Findings

Achieves state-of-the-art forecasting accuracy

Effectively models physical dependencies among cyclone attributes

Outperforms existing deep learning methods on multiple datasets

Abstract

Tropical cyclone (TC) forecasting is critical for disaster warning and emergency response. Deep learning methods address computational challenges but often neglect physical relationships between TC attributes, resulting in predictions lacking physical consistency. To address this, we propose Phys-Diff, a physics-inspired latent diffusion model that disentangles latent features into task-specific components (trajectory, pressure, wind speed) and employs cross-task attention to introduce prior physics-inspired inductive biases, thereby embedding physically consistent dependencies among TC attributes. Phys-Diff integrates multimodal data including historical cyclone attributes, ERA5 reanalysis data, and FengWu forecast fields via a Transformer encoder-decoder architecture, further enhancing forecasting performance. Experiments demonstrate state-of-the-art performance on global and regional datasets.