Bayesian optimized random forest models for predicting permanent displacement considering fault types and pulse-like ground motion effects

Earthquake-induced landslides can cause severe casualties and substantial economic losses, making seismic slope stability a crucial issue in geotechnical earthquake engineering. Although permanent displacement has been widely used to evaluate slope response during earthquakes, many existing predictive displacement models (PDMs) do not explicitly account for fault type, ground motion orientation, or the probability of pulse-like ground motion (PLGM). This study quantifies how these factors influence permanent displacement. The results show clear fault-mechanism-dependent directional effects: for non-strike-slip faults, larger displacements are more likely when motions are oriented approximately fault-normal, whereas for strike-slip events, larger displacements tend to occur in directions closer to fault-parallel. Additionally, higher PLGM probability is associated with both larger permanent displacement and stronger orientation sensitivity. To capture these relationships, new PDMs are developed using a Bayesian-optimized random forest regressor that explicitly incorporates pulse probability and slope-fault orientation, enabling direction-dependent displacement estimates within a unified framework. A regional case-study application to the 2017 Jiuzhaigou earthquake, using an independent landslide inventory, shows a clear displacement-based hazard gradient, with mapped landslides showing a progressive enrichment in areas with higher predicted-displacement classes. These results indicate that the proposed models provide a practical basis for regional seismic landslide hazard categorization and screening. The models calibrated for Mw 5.0–8.0 earthquakes within 120 km of the fault, and their applicability to weaker shaking and different site conditions still requires further validation.