Application of Hybrid Ensemble Machine Learning Model for Peak Ground Acceleration Prediction: Case Studies on Regional Earthquakes

ABSTRACT This study uses different nonlinear machine learning (ML) models and a hybrid ensemble surrogate model to predict peak ground acceleration (PGA), an important parameter in earthquake ground motion. Typically, models with established equations based on linear regression are utilized to create a ground motion model (GMM); however, traditional ground motion models based on linear regression often struggle to capture the inherently complex and nonlinear behavior of earthquake ground motion. Hence, this paper applies various nonlinear machine learning (ML) models and a hybrid stacked ensemble model (HySEM) to achieve better prediction of the ground motion. For the study, more than 6000 records were compiled from the PEER ground motion database, focusing on Shallow Crustal Earthquakes in Active Tectonic Regions. The variables included in the dataset are, namely, 5%–95% Duration (sec), Magnitude, Epicenteral Distance (km), Hypocentral Distance (km), Shear wave velocity, V s 30 (m/s), Peak ground acceleration ( g ). It was observed that HySEM exhibited the best results with R 2 equal to 0.955. Furthermore, the DALEX framework was applied to HySEM, the best‐performing model, to gain insights into feature importance and interactions. To check the model's adaptability, HySEM was adopted to predict PGA for historic seismic events in Nepal and Japan, which were not a part of the training data. A comparison of these predictions with existing attenuation laws for those regions demonstrated that HySEM could predict PGA with significant accuracy in various scenarios. This research underscores the potential of ML techniques for developing efficient and accurate region‐specific ground motion prediction capabilities, offering an alternative, especially in areas with high seismic activity but limited recorded data.