Evaluating Regionally Diverse Pathways to Unprecedented Heat Extremes Across Africa in NeuralGCM

Abstract Extreme heat is among the deadliest hazards facing Africa, a continent where reliable forecasting is most needed yet hardest to provide. Hybrid atmospheric models, which couple physics-based dynamical cores to learned physics, now rival conventional prediction at a fraction of the computational cost, yet their learned components carry no guarantee on the rare, record-breaking events that matter most for heat risk. Whether such a model reaches unprecedented extremes through the correct physical processes remains untested. We evaluate NeuralGCM, a hybrid model whose differentiability we use as an instrument for validation. After confirming that NeuralGCM reproduces four climatically distinct African heat events against ERA5 reanalysis, we drive each event towards its plausible worst case by gradient-based optimization of the initial state and ask whether the amplification proceeds through the mechanism established for that region. The optimized events intensify through distinct, regionally appropriate pathways, with HI amplifications of 1.2 to 3.0 ◦C: thermally driven under mid-tropospheric ridging and subsidence in the dry regimes of southern and northeastern Africa, moisture driven under weaker circulation in humid West Africa, and intermediate in East Africa. NeuralGCM thus passes a regionally differentiated process test, demonstrating that differentiability itself provides the instrument for validating whether hybrid models intensify heat extremes through genuine physics.