An Interpretable Machine Learning Approach in Understanding Lateral Spreading Case Histories

Abstract

Lateral spreading is one of the most common secondary earthquake effects that cause severe damage to structures and lifelines. While there is no widely accepted approach to predicting lateral spread displacements, challenges to the existing empirical and machine learning models include obscurity, overfitting, and reluctance of practical users. This study reveals patterns in the available lateral displacement database, identifying rules that describe the significant relationships among various attributes that led to lateral spreading. Seven conditional attributes (earthquake magnitude, epicentral distance, maximum acceleration, fines content, mean grain size, thickness of liquefiable layer, and free-face ratio) and one decision attribute (horizontal displacement) were considered in modeling a binary class rough set machine learning. There are eighteen rules generated in the form of if-then statements. The decision support system reveals that the severity of lateral spreading clearly comes from the combinations of relevant attributes. Moreover, five clusters of rules were also observed from the generated rules. Useful information regarding the different lateral spreading case scenarios emerges from the results. Statistical validation and interpretation of rules using principles of soil mechanics and related studies were also performed. The output of this study, a decision support system, can be very useful to decision-makers and planners in understanding the lateral spreading phenomena. Recommendations for the model improvement and for further studies were discussed.