Physically constrained machine learning improves radiation, clouds, and albedo forecasts globally

Abstract

Clouds, radiation, and surface albedo are essential to Earth’s energy balance and have a significant impact on applications such as solar energy, yet they remain challenging to represent accurately in global forecast systems. Recent machine learning-based weather forecasting models have matched or outperformed conventional numerical models for large-scale dynamical variables, but most do not explicitly predict clouds, radiation, or albedo, limiting their physical fidelity. Here, we present FuXi-RTM, a hybrid machine learning model that integrates a differentiable machine learning surrogate of a widely used radiative transfer model (MLRTM) directly within global weather forecasts. By jointly predicting clouds, surface albedo, and radiative fluxes, and by allowing radiative transfer gradients to constrain the forecast state, FuXi-RTM improves radiative consistency while maintaining computational efficiency. Against reanalysis data, independent ground-based station observations, and satellite-derived radiation products, FuXi-RTM improves forecasts of shortwave radiation and cloud-related variables relative to unconstrained variants and state-of-the-art physics-based models. These results suggest that differentiable physical constraints can improve machine learning-based weather forecasts for both Earth system prediction and solar energy applications.