Ensemble machine learning models for predictive analysis: Application to seismic ground motion data

Computer science has become an essential discipline for solving complex, data-intensive problems across the natural sciences. This study demonstrates how machine learning algorithms—especially ensemble methods such as stacking, random forest, and gradient boosting—can be used to build data-driven ground-motion models (GMMs) for predicting peak ground acceleration (PGA), a key parameter in seismic hazard assessment. The stacking approach integrates multiple base learners (linear regression, polynomial regression, decision tree, and random forest) with meta-models (linear regression, decision tree, or random forest) to enhance prediction accuracy. Random forest constructs an ensemble of decision trees, while gradient boosting sequentially refines residuals to minimize errors. Models are trained on over 10000 records from small-to-moderate earthquakes (Mw 3.5–5.8) with hypocentral distances up to 200 km. Predictor variables include moment magnitude (Mw), hypocentral distance (Hypo-D), rupture-top depth (Ztor), and average shear-wave velocity in the upper 30 m (VS30). Performance evaluation reveals that the stacked model with a linear-regression meta-model achieves the highest accuracy, underscoring the potential of ensemble learning for seismic hazard modeling.