Peak Ground Acceleration Prediction - Regression Based Machine Learning Models

Volume Title: ICASGE2025

Paper ID : 1171-ICASGE

University of Memphis, TN, Memphis, USA

One of the most critical aspects of civil engineering and structural design is accounting for seismic hazard to ensure the safety and resilience of infrastructure. A key parameter in seismic hazard analysis is the peak ground acceleration (PGA), which represents the maximum acceleration experienced by a site or structure during an earthquake. Ground-motion models (GMMs) serve as essential tools in capturing the complex relationships between ground motion intensity measures and various predictor variables. This study develops data-driven GMMs using both parametric and nonparametric machine learning algorithms, including linear regression, polynomial regression, decision tree, and random forest. The models are trained on an extensive database containing over 10,000 ground-motion records from small-to-moderate earthquakes with magnitudes ranging from 3.5 to 5.8 and hypocentral distances up to 200 km. Predictor variables include moment magnitude (Mw), hypocentral distance, VS30, and focal depth. By evaluating model performance through comparisons of predicted and observed PGA values, results indicate that the random forest model outperforms traditional regression-based GMMs. The findings highlight the suitability of machine learning-based GMMs, particularly random forest, for seismic hazard applications in regions characterized by frequent small-to-moderate earthquakes.

Peak Ground Acceleration; Earthquake; Machine Learning; Regression Model

Geotechnical Engineering (GEO)