Abstract: In order to establish a prediction method for the ground peak acceleration (PGA) that does not depend on the soil constitutive model but only on the ground motion and site characteristics, six characteristic parameters are chosen through the feature selection based on 3104 groups of bedrock and surface seismic records collected from the KiK-net strong-motion seismograph network of Japan. Then, the input ground motion characteristics are characterized through the peak bedrock acceleration and predominant frequency, and the site characteristics are characterized by the soil depth at shear wave velocity of 800 m/s, site fundamental period, bedrock shear wave velocity and surface shear wave velocity. The XGBoost model in machine learning is used to establish the prediction models for the PGA based on the above six characteristics. It is shown that the prediction results of the XGBoost prediction model are stable and can be used to predict the PGA better by comparing the records and one-dimensional numerical simulation methods. The coefficients of determination of the training set and the test set are greater than 0.925, and the mean absolute percentage errors are about 20%, which is obviously better than the one-dimensional numerical simulation methods. At the same time, the SHAP is introduced to analyze the influence and dependence between the input characteristics and the predicted results, which enhances the interpretability of the model and provides reliability support for the predicted results.