Interpretable machine learning framework for performance-based retrofit scheme of blast-damaged reinforced concrete columns

Abstract

Explainable artificial intelligence (xAI) has been widely used to improve learning performance because it helps users understand the learning processes. This paper proposes an xAI-based framework to build retrofit schemes for blast-damaged RC columns. This framework includes a multi-stage learner rapidly predicting blast resistance levels using simple structural details. The extensive data for the blast resistance was analyzed with a three-step interpreting process: (1) partial dependence plot (PDP) to initially judge whether the retrofit is effective, (2) 1D accumulated local effect (ALE) to set the quantitative retrofit thresholds for ductility- and stiffness-related variables, and (3) 2D ALE to build effective retrofit schemes considering the interactive effects of retrofit variables on blast resistance. Based on the interpretation results, the various retrofit schemes were recommended for the column failure types and expected damage conditions. Overall, multiple retrofit schemes were required for the columns to accommodate the expected severe and moderate damage conditions.