Machine learning based site amplification prediction models for shallow bedrock sites

Abstract

Estimation of site amplification is important for a seismic hazard assessment. To develop site amplification prediction models, the models have utilized both measured and randomized profiles to compensate for the limited number of measurement database. One-dimensional site response analyses were performed to calculate linear and nonlinear responses, which were then separated into training and test data. While most amplification models are based on regression analysis using specific types of functions, this study utilized machine learning (ML) models with algorithms independent of prior functional forms. Random forest (RF) and deep neural network (DNN) models were used to train the model, and the test data were used to predict the amplifications. The DNN-based model shows more accurate results than the RF-based model, and the ML-based models were shown to outperform the existing regression-based model