AI and the World's Hard Problems · Earth & climate science · arXiv · 2026-10-21
Improved Precipitation Variability in the Community Atmosphere Model through Implementation of a Neural-Network Deep Convection Scheme
Deep convection schemes remain a primary source of uncertainty in global climate models, and replacing them with data-driven alternatives while maintaining physical consistency and numerical stability is an open challenge. The authors present a hybrid CAM6 configuration in which a fully connected deep neural network, trained on coarse-grained output from the storm-resolving global model gSAM to derive subgrid-scale convective fluxes and tendencies, replaces the Zhang-McFarlane deep convection scheme while radiation, turbulence, and other cloud processes stay unchanged. The network runs stably for more than a decade while conserving energy and moisture, produces a broadly realistic climatology without retuning the remaining parameterizations, markedly reduces light- and extreme-precipitation biases, and shifts the diurnal cycle of tropical land precipitation from a spurious midday peak to the observed late-afternoon peak.