PNNL

Abstract

A critical challenge for machine-learning (ML) parameterizations in global climate models (GCMs) is to achieve stable, accurate simulations under climates not seen during training. Previous studies have demonstrated promising offline performance and year-long online stability in aquaplanet simulations but have encountered difficulties in real geography and under climate warming. Here we report that a GCM with real geography configuration using neural-network-based cloud and convection parameterization, trained exclusively with the present-day climate data, successfully performs a stable, decade-long simulation of a warm climate with +4K sea surface temperature (SST). The neural network is based on Han et al. (2023), revised by including additional inputs for the top-of-model solar radiation, surface-emitted longwave radiation, and land surface fraction. The decadal online simulation captures the global precipitation distribution, surface air temperatures, vertical atmospheric structures, and extreme precipitation very well, closely matching simulations from the superparameterized CAM (SPCAM) in the warm climate without accuracy degradation compared to those in the baseline climate. Moreover, it produces a climate response to +4K SST in atmospheric thermodynamic states and circulations similar to those from SPCAM and the default CAM5. To our knowledge, this is the first time an ML parameterization successfully achieves online extrapolation to a warm climate without using additional warm-climate data for training. This work demonstrates the potential of ML-driven parameterizations for credible long-term climate projections.