Boosting Extreme Weather Prediction by Fine-tuning a Pre-Trained Large Model: A Study on GraphCast

Large-scale deep learning weather prediction models are revolutionizing the field of weather forecasting. GraphCast is the current state-of-the-art model, but its training is not explicitly designed for predicting extreme weather events. However, extreme weather prediction is more critical because it directly impacts public safety, potentially saving a lot of lives and resources. This paper improves large-scale weather models such as GraphCast by introducing an uncertainty estimation module to differentiate the importance of extreme weather data. We hypothesize and demonstrate that regions with higher uncertainty are more prone to cause prediction errors. By fine-tuning large-scale weather prediction models such as GraphCast with our uncertainty-aware weighting method, we enhance extreme weather forecasting in extreme cases where predictions were previously poor. Our approach provides a pathway for more accurate extreme weather forecasts and a pipeline for future model fine-tuning efforts.