Improving our understanding of the impact of extreme climate events on crop yields is critical to increasing the resilience of food systems. This study aimed to explore the impact of extreme climate events at different growth stages on maize yields by incorporating a crop model with machine learning methods, and considering the spatial heterogeneity of climatic conditions in a typical black soil region. A method that couples SKATER, APSIM-NGMaize and machine learning was proposed and used to estimate regional dominant extreme climate factors on spring maize yield. The increase in the mean squared error method and sliding window-based correlation coefficients were used to describe the non-linear response dynamics and threshold-dependent regime shifts of the impact of extreme climate indices on maize yield. The RF-APSIM model better explained historical crop yield variations (R2 = 17.8%-48.1%) compared to LASSO-APSIM and MLR-APSIM, although predictive skill remained inherently limited by agricultural data noise. For the semiarid plain zone, maize yield was mainly affected by the combination of excess precipitation and extreme heat, especially after the floral initiation stage. The mountainous zone exhibited strong resistance to waterlogging because its topography facilitates drainage, with extreme low temperatures being the main limiting factor. The results highlight that spatial partitioning is necessary even at the provincial scale due to zonal differences in yield responses. For improving resilience, heat and flood tolerance traits are important in plain zones, while cold tolerance traits are more important in mountainous areas. These findings provide a methodology for unravelling the non-linear mechanisms of extreme climate impacts on crops.