New Perspectives in Causal Relationships Between the Response of a Rocking Block and Intensity Measures via Ensemble Machine Learning Methodologies

This paper investigates the ability of machine learning (ML) to characterise the response of rocking structures when subjected to recorded earthquakes. In particular, it uses the structural parameters of a rigid block and strong ground motion characteristics to train two random forest (RF) models. The first model predicts whether a block, given that it initiates rocking motion, overturns or undergoes safe rocking, and identifies the main variables, i.e., structural and ground motion features, that govern such classification. Provided no overturning occurs, the second RF model predicts the peak rocking rotation of a block under ground motion records. Importantly, this study also employs interpretable ML techniques (such as partial dependence plots and SHAP additive explanations) to identify causal relationships between strong ground motion parameters and rocking response. The analysis shows that under high-intensity earthquakes, the peak ground velocity (PGV) governs the overturning of a rocking block. For earthquakes of moderate intensity, overturning becomes a more interactive phenomenon where the PGV, frequency/period and duration characteristics of the seismic signal contribute. Finally, this research shows that high safe rocking amplitude is also interactive, with velocity, displacement, (mean) frequency/period, and duration characteristics of the ground excitation playing a pivotal role.