Development of a site and motion proxy-based site amplification model for shallow bedrock profiles using machine learning

Duhee Park¹Yonggook Lee^1*Oh-Sung Kwon²

• ¹Hanyang University, Seoul, Republic of Korea
• ²University of Toronto, Toronto, Ontario, Canada

We developed proxy-based linear and nonlinear site amplification models for shallow bedrock sites using machine learning (ML). The outputs of a series of one-dimensional site response analyses were used for training. Three ML algorithms were used: random forest (RF), extreme gradient boosting (XGB), and deep neural network (DNN). For the training, four sites and two motion proxies were used. The accuracy of the trained model was compared against the results with a conventional regression-based equation model. In addition, the proxy-based ML model was also compared against a rigorous ML model that utilized the entire shear-wave velocity profile and input motion response spectrum as input data. When identical proxies were used, the differences between the regression and ML-based models were not pronounced. However, when the ML model was trained simultaneously with the site and motion proxies for both linear and nonlinear components, the prediction performance was significantly enhanced. This revealed that the traditional two-track approach of the site-proxy-dependent linear component and motion-proxy-conditioned nonlinear component is ineffective. We recommend a pairing scheme for site and motion proxies that produces the most accurate predictions. Among the three ML methods, the RF algorithm exhibited the weakest performance. The XGB and DNN algorithms' prediction accuracies were superior to the RF algorithm. The XGB and DNN outperformed each other when predicting the linear and nonlinear components, respectively. In quantitative terms, the proposed ML models achieved coefficient of determination (R 2 ) values up to 0.97 with root mean square error (RMSE) as low as 0.04 for linear components, and R 2 up to 0.92 with RMSE as low as 0.06 for nonlinear components, demonstrating significant improvements over conventional regression-based models. Compared with a rigorous ML model, the proxy-based models exhibited agreeable predictions with far less information, illustrating the benefit of adopting the ML algorithms for improved adaptability and predictive capability. The constraint imposed on the site type, considering only profiles with a bedrock depth of less than 30 m, may have resulted in the strong performance of the proxy-based model. Further studies are required to explore the ability of proxy-based ML model to predict site amplification.