Physics-inspired machine learning for earthquake-induced landslide susceptibility mapping: integrating fault-controlled seismic energy attenuation and directional effects

Most susceptibility models fail to capture the coupled effects of fault-controlled seismic attenuation and directional wave propagation, limiting their ability to represent spatial heterogeneity. This research proposes a Fault Azimuth Attenuation Coupling Factor (FAACF) to quantitatively characterise the synergistic ‘direction–distance’ effects of seismic dynamics. Fourteen conditioning factors were selected from five categories: topography, seismic activity, geological conditions, vegetation cover, and human activity. Together with the newly proposed FAACF, these yielded a total of 15 factors for landslide susceptibility modelling. The results show that landslides predominantly concentrated in areas characterised by strong positive (>0.17) or strong negative (<−0.16) FAACF values. Incorporating FAACF significantly improves model performance, increasing the AUC to 0.926 for the XGBoost model and to 0.915 for the RF model. SHAP-based interpretability analysis identifies elevation, FAACF, slope, and peak ground acceleration (PGA) as the dominant controlling factors and reveals clear threshold effects. In addition, strong positive interactions are observed between elevation and PGA, as well as between FAACF and elevation, jointly regulating slope instability. This study highlights the critical role of seismic dynamic characteristics in landslide susceptibility modelling and demonstrates that the proposed FAACF provides a transferable reference for assessing earthquake-induced landslide susceptibility in other tectonically active regions.