Prediction of damage potential in mainshock-aftershock sequences using machine learning algorithms

Assessing the potential damage caused by earthquakes is crucial for a community's emergency response. In this study, four machine learning (ML) methods-random forest, extremely randomized trees, AdaBoost (AB), and gradient boosting (GB)-were employed to develop prediction models for the damage potential of the mainshock (DIMS) and mainshock-aftershock sequences (DIMA). Building structures were modeled using eight single-degree-of-freedom (SDOF) systems with different hysteretic rules. A set of 662 recorded mainshock-aftershock (MS-AS) ground motions was selected from the PEER database. Seven intensity measures (IMs) were chosen to represent the characteristics of the mainshock and aftershock. The results revealed that the selected ML methods can well predict the structural damage potential of the SDOF systems, except for the AB method. The GB model exhibited the best performance, making it the recommended choice for predicting DIMS and DIMA among the four ML models. Additionally, the impact of input variables in the prediction was investigated using the shapley additive explanations (SHAP) method. The high-correlation variables were sensitive to the structural period (T). At T=1.0 s, the mainshock peak ground velocity (PGVM) and aftershock peak ground displacement (PGDA) significantly influenced the prediction of DIMA. When T increased to 5.0 s, the primary high-correlation factor of the mainshock IMs changed from PGVM to the mainshock peak ground displacement (PGDM); however, the high-correlation variable of the aftershock IMs remained PGDA. The high-correlation factors for DIMS showed trends similar to those of DIMA. Finally, a table summarizing the first and second high-correlation variables for predicting DIMS and DIMA were provided, offering a valuable reference for parameter selection in seismic damage prediction for mainshock-aftershock sequences.