Abstract Heat extremes have become a major health hazard around the world. Understanding their mechanisms remains a major challenge because the physical drivers interact in a nonlinear way. Here we introduce a globally perturbed reforecast framework driven by the Neural general circulation model (NeuralGCM). Sensitivity reforecast experiments that independently remove initial condition anomalies over spatially distinct patches identified the high impact regions (HIRs) for the record‐breaking August 2022 South China heatwave (SCH22) in Europe and North America (NA) through changes in forecast skill, which are further confirmed by dynamic diagnostics. Forecasts initialized using anomalies only from HIRs covering just 25% of the global domain successfully reproduce the evolution and spatial pattern of SCH22. These findings can also generalize to another AI‐based weather model FuXi. Our proposed framework helps to improve accessibility to global‐scale diagnostic for extreme events with robust results.
Xiang et al. (Sat,) studied this question.