Conditional generative adversarial networks for the generation of strong ground motion parameters using KiK-net ground motion records

Abstract

Prediction of ground motion parameters is important for seismic risk analysis and the seismic design of structures. The limited availability of strong ground motion records (peak ground acceleration (PGA) > 100 gal) limits the application of machine learning techniques in this domain. In this study, we utilized the KiK-net database to identify three site parameters—V_S30, V_S-Zhole, and borehole depth—corresponding to the peak underground acceleration (PUA) on the downhole and spectral acceleration (Sa) across 15 different periods as input features for the machine learning model. The output parameters were PGA and Sa across 15 different periods. Pairs of strong ground motion records (PGA > 100 gal) are extracted from the original KiK-net database and processed using a conditional generative adversarial network (cGAN) model to synthesize ground motion parameters, then the original KiK-net database is supplemented. Regression prediction tests were performed on both the original and augmented databases and evaluated using three regression metrics: root mean square error (RMSE), mean absolute error (MAE), and R². The findings indicate that the model trained with synthetic data generated by the cGAN enhances the prediction accuracy of ground motion parameters. A comparative analysis of the ground motion amplification coefficients between station records and cGAN-generated data across various spectral periods was performed. The amplification coefficients across different spectral periods aligned with the ground-motion records, thereby validating the reliability of the cGAN-generated synthetic ground-motion data from a seismological perspective. Furthermore, the ground motion parameters produced by the cGAN represent a valuable resource for future machine learning investigations.