Machine-learning interpretability techniques for seismic performance assessment of infrastructure systems

Abstract Machine-learning has recently gained considerable attention in the earthquake engineering community, as it can map the complex relationship between the expected damage and the input parameters. It is often necessary to understand the reasons for the behavior and predictions of the machine-learning model. This paper addresses this issue through interpretable machine-learning approaches such as partial dependence plots, accumulated local effects, and Shapely additive explanations. The evaluation of these approaches is carried out (1) at a component level by analyzing the shear strength predictions by a machine-learning model and (2) at a regional level through the machine-learning model for the regional damage assessment of bridges in California. The comparison helps to identify (1) the proper implementation of these approaches for the efficient use of machine-learning models and (2) key influential variables and thresholds that govern the prediction of the machine-learning models.