Machine Learning-Based Assessment of the Seismic Performance of Geotechnical Systems

抄録

In assessing the seismic performance of geotechnical systems, engineers often use analytical models to estimate the amount of seismically-induced displacements. These models consider system properties, earthquake parameters, and ground motion intensity measures (IMs) as inputs and have been typically formulated using “traditional” statistical techniques. This study discusses a new set of machine learning (ML) based models to estimate seismically-induced slope displacements in subduction and shallow crustal earthquake zones and liquefaction-induced building settlements in shallow crustal earthquake zones. Machine learning is used to select efficient features that explain seismically-induced displacements and settlements. The feature selection suggests no significant gain in accuracy beyond a small subset of features. Based on the selected features, a set models is developed considering several ML-based techniques with varying flexibility, interpretability, and bias-variance trade-offs. The developed models are assessed by evaluating test errors, their scaling compared to existing models, and their performance in case histories. The developed ML-based models contribute to performance-based assessments and also enhance the treatment of epistemic uncertainties in estimating seismically-induced displacements. Lastly, as discussed in the paper, caution should be exercised when assessing ML models as they could be perceived as a “black box” without proper context and extrapolate inappropriately.