Machine learning-based models for estimating seismically-induced slope displacements in subduction earthquake zones

Abstract

The assessment of the seismic performance of slope systems often relies on estimating the amount of seismically-induced slope displacements (D) using semi-empirical D models. These models take as inputs the slope properties and ground motion intensity measures (IMs) to provide D estimates that are used in engineering design. However, most of the available D models have been developed for regions affected by shallow crustal seismicity. Comparatively, the available D models for subduction tectonic settings are scarce. Moreover, most existing models have been developed using traditional statistical methods that do not take advantage of modern data-driven approaches. In this study, we develop new machine learning (ML) based D models applicable to subduction earthquake zones (considering both interface and intraslab mechanisms) using the NGA-Sub ground motion database. A systematic feature selection is performed using three ML procedures, finding that the yield coefficient (ky), the initial fundamental period of the slope system (Ts), the earthquake magnitude (M), the peak ground velocity (PGV), and the pseudo-spectral acceleration at 1.3Ts(Sa(1.3Ts)) are efficient features for estimating D in subduction earthquake zones. Based on the selected features, we develop five ML-based D models by using modern ML procedures (i.e., ridge regression, random forest, gradient boosting decision tree (GBDT), support vector regression (SVR), and residual neural network (ResNet)). The developed ML-based D models do not need to be restricted to predefined fixed functional forms, as has often been the case for previously developed D models. The ridge regression model is used to represent generic traditional D models as it has a polynomial-based functional form. Compared to the ridge regression model, the other developed ML-based models are able to better capture the complex relationship between D and the slope properties and IMs, showing a better predictive performance on test tests; hence, they outperform traditional models. In addition, we compare the performance of the developed ML-based models in terms of predictive performance, model trends, and computational cost for training. Lastly, the developed ML-based models also enhance the treatment of epistemic uncertainty in the estimation of D, given the scarcity of available robust D models for subduction zone tectonic settings.