Machine learning based ground motion site amplification prediction

Abstract

Site condition impact on seismic ground motion has been a complex but important subject in earthquake hazard analysis. Traditional studies on site amplification effect are either based on site response via wave propagation simulation or regression analysis using parameters such as Vs30, bedrock ground motion and site response period. Ground Motion Prediction Equations (GMPEs) are used for regions where there is limited data of seismic records. The main issues with these approaches are that they cannot demonstrate the complex relationship between site amplification and its various affecting parameters, thus there exists large uncertainty in the results. Recent studies based on machine learning have shown significant improvement in predicting the site amplification, but the result is not well explained. This study assembled the information on 6 parameters including Vs30, magnitude, epicentral distance, earthquake source depth, bedrock ground motion, and altitude of 353,327 records observed during 1997 and 2019 from 698 KiK-net stations. Three machine learning algorithms of Random Forest (RF), XGBoost, and Deep Neural Networks (DNN) were implemented to predict the site amplification factor using these 6 selected parameters. Shapley Additive explanation (SHAP) was used to explain the importance of the 6 parameters. The results show that all three machine learning algorithms performed much better than the traditional GMPE approach with XGBoost’s performance the best. The explanation provided by the SHAP analysis further enhanced the reasonability of this study. It is anticipated that the combination of machine learning and SHAP analysis can provide better assessment for site amplification of ground motion with better potential of future application in seismic hazard analysis.