Prediction Method of 3D Ground Motion Parameters for Broadband Scenario Earthquake Based on Deterministic Physical Simulation and Machine Learning

In order to incorporate the properties of surface low-wave velocity soils (heterogeneity,nonlinearity,filter amplification) into the physics-based simulation and prediction of 3D ground motion parameters,this paper proposes a hybrid method combining physics-based ground motion simulation and Random Forest model.The paper describes the process for predicting ground vibration parameters in three directions (East-West,North-South,and Vertical) using data from the KiK-Net strong motion database.Firstly,the ground motion parameters are predicted using the Random Forest model trained on site parameters and ground motion parameters from selected bedrock surfaces.Then,the physics-based ground motion simulation method is used to obtain the ground motion parameters at the bedrock surfaces.Finally,the trained Random Forest model is used to predict the 3D ground motion parameters using the simulation results and site parameters as inputs.The trained Random Forest model predicts the ground vibration parameters at different cycles of the ground surface using the simulation results and site parameters as inputs.It outputs the peak acceleration (PGA) and spectral acceleration (Sa) at the ground surface location.After evaluating the prediction results of the Random Forest model,taking the 2016 M_w 6.2 Tottori earthquake in Japan as an example,the broadband (0.1~10.0Hz) results of the method are compared with the observed data and ground motion prediction equation to verify the effectiveness and practicability of the proposed method.The results of the ground motion parameters distribution in the region are presented based on the method.