Weather Modification for Heavy-Rainfall Mitigation

Data-driven weather forecasting with deep learning has advanced rapidly in recent years. We develop representation-learning methods that handle many weather variables efficiently, achieving accurate forecasts with far fewer computational resources than conventional approaches.

We’re also pursuing a more ambitious direction: intervening on weather fields themselves, in a physically plausible way, to mitigate damage from extreme events such as heavy rainfall. Unlike adversarial perturbations, our approach guides the sampling trajectory of diffusion-based weather models so that any intervention remains consistent with physical law.

Progress so far

We presented VarteX, an efficient weather-variable representation learning model, at an ICML 2024 workshop. We’ve also proposed a method for physically plausible precipitation-mitigating interventions in diffusion-based weather models, more grounded than adversarial perturbations. This line of work is also supported by JST’s Moonshot R&D program (AMAGOI project).