An Interpretable Weather Forecasting Model With Separately-Learned Dynamics and Physics Neural Networks

Machine learning (ML) offers a promising alternative for weather forecasting by reducing computational costs and modeling complex non-linear atmospheric processes. While recent foundation models highlight this potential with advanced architectures, interpreting the “black-box” nature of ML models remains challenging. This study proposes an interpretable ML model combining graph neural networks and multi-layer perceptrons (MLP). By using the graph targeted for large-scale movement in the dynamical core, and MLP targeted for small-scale motion in physical parameterizations, our model provides a new perspective to simulate the transition of variables. Through 10-day iterative forecasts, our model shows comparable performance to purely data-driven models when trained at 1.5° resolution, with fewer parameters, and faster training speed than physics-informed neural networks, like those solving differential equations. Moreover, a case study of the 2020 monsoon demonstrates the model's interpretability by exploring the correlations between the attentions in graphs and atmospheric processes such as wind and precipitation.