Development of machine learning-based site amplification models for Japan from borehole recordings

Abstract

A large number of site amplification models have been developed using regression and machine learning (ML) approaches. Although ML models generally outperform traditional methods in predicting site amplification, the effects of specific site and motion parameters on model accuracy remain insufficiently explored. Using a meta-dataset of earthquake recordings from Japan’s Kiban Kyoshin Network (KiK-net), we trained six ML-based site amplification models: random forest (RF), extreme gradient boosting (XGB), and deep neural network (DNN), together with their hybrid variants incorporating Bayesian optimization (BO). A sensitivity analysis using RF evaluated how combinations of input proxies influence predictive performance, leading to the identification of an optimal proxy configuration. Among the six models trained with this configuration, BO-DNN performed best at period T < 0.1 s, whereas BO-XGB showed superior performance at T > 0.1 s. Shapley Additive exPlanations (SHAP) was used to rank proxy importance, identifying the peak frequency of the horizontal-to-vertical spectral-ratio curve (fp), the time-averaged shear-wave velocity up to 30 m (Vs30), borehole depth (BD), and borehole spectral acceleration averaged over 0.1–0.3 s (SS) as the most influential proxies. The proposed models demonstrate superior performance compared with two previously published models that were also developed using KiK-net data.