<p>Marine heatwaves (MHWs) are extreme ocean warming events that profoundly impact marine ecosystems, yet the relative contributions of atmospheric and oceanic drivers before MHW onset in the South China Sea (SCS) remain unclear. To address this, we divided the SCS into five subregions and applied long short-term memory (LSTM) models integrated with the expected gradients method to assess the changing contributions of six key drivers leading to MHW events. Our results show that mixed layer depth is the dominant contributor on the day before MHW onset, reflecting its key role in modulating the ocean’s thermal structure. In contrast, the influence of 10-meter wind fields shows greater spatiotemporal fluctuations compared to that of downward solar shortwave radiation and surface geostrophic currents. Clustering analysis identified five distinct contribution patterns, highlighting pronounced regional differences in MHW-driving mechanisms. The effective contribution windows of each factor are shaped by wind field characteristics, mixed layer dynamics, and local topography. In open-sea regions, zonal winds exert an earlier and more sustained influence, highlighting their role in facilitating horizontal heat transport. Conversely, their impact in nearshore areas is more delayed and short-lived, likely due to modulation by complex coastal topography that affects local circulation and vertical mixing. Meanwhile, shallower mixed layers in these coastal regions tend to exhibit more prolonged contributions, as they respond more rapidly to surface heat fluxes. Additionally, meridional wind plays a crucial role across all regions from 30 to 5 days before MHW onset, marking a key window for MHW prediction. This study demonstrates the effectiveness of interpretable deep learning in uncovering MHW onset mechanisms and provides a foundation for developing region-specific predictive models for the SCS.</p>