Global Climate Modeling with Improved Precipitation Characteristics by Learning Physics (GRIST-MPS v1.0) from Global Storm-Resolving Modeling

This study develops a machine learning (ML)-based physics parameterization suite trained on 80-day global storm-resolving model (GSRM) simulation data, attempting to replace all conventional physics tendencies in a general circulation model (GCM). Our approach strategically selects key prognostic variables as input features, enabling an effective emulation of multiscale flow interactions of the GSRM by the GCM via dynamics-physics coupling. The resulting ML-enhanced GCM achieves stable Atmospheric Model Intercomparison Project (AMIP)-type simulations over six years, surpassing its conventional counterpart with improved precipitation performance—reducing root-mean-square errors by 8 % in boreal summer and 16 % in winter, compared to observations. Moreover, the hybrid ML-GCM better captures precipitation frequency–intensity spectra, notably mitigating the overproduction of light tropical rainfall and improving the simulation of moderate rain rates. Sensitivity experiments using different neural network architectures (ResNet, CNN, DNN) demonstrate that all configurations can maintain long-term simulation stability, with ResNet showing superior capability in the simulation accuracy. This work presents a transferable framework that leverages km-scale GSRM data to enhance GCM performance via ML integration, offering a potential route to reduce the gaps between two modeling paradigms.