XGBoost for Predicting Peak Ground Acceleration Based on the NGA West2 Ground Motion Database

Abstract

Earthquakes remain among the most destructive natural hazards worldwide, causing significant loss of life, economic damage, and infrastructure failures each year. Accurate prediction of peak ground acceleration (PGA), a key indicator of ground motion intensity, is fundamental to reliable seismic hazard assessment and earthquake-resistant structural design. Inaccurate or oversimplified PGA estimates can lead to unsafe structural designs or overly conservative construction costs, directly affecting public safety, infrastructure resilience, and economic sustainability. However, conventional ground motion prediction equations are typically based on predefined functional forms that may not fully capture the complex, nonlinear relationships among seismic source, path, and site parameters, leading to inaccuracies in PGA estimation. To address this limitation, this study develops an extreme gradient boosting (XGBoost) regression model to predict PGA using the NGA-West2 ground motion dataset from the Pacific Earthquake Engineering Research Center (PEER). The dataset, filtered to 16,211 records, includes moment magnitude (Mw), rake angle (RA), hypocentral depth, closest rupture distance (ClstD), and average shear-wave velocity (Vs30) as input features. At the same time, the logarithm of PGA serves as the target variable. The model achieved high accuracy on both training (R² = 0.971, RMSE = 0.184) and testing (R² = 0.936, RMSE = 0.269) datasets, demonstrating excellent learning and generalization performance. Feature importance analysis identified Mw and ClstD as the dominant predictors of PGA, accounting for approximately 64% and 29% of the total contribution, respectively. The results confirm that the XGBoost model effectively models nonlinear dependencies and complex interactions among seismic parameters.