Two-stage prediction of drift ratio limits of corroded RC columns based on interpretable machine learning methods

RC columns exposed to harsh environments are susceptible to internal reinforcement corrosion, leading to a reduction in lateral deformation capacity. The accurate prediction of drift ratio limits (DRLs) for corroded RC columns (CRCCs) across various damage states is crucial for reliable damage assessment and seismic resilience analysis. Current literatures remain inadequate for predicting DRLs for CRCCs with diverse service life. To address this gap, this paper introduces a two-stage machine learning (ML) approach for the simultaneous prediction of DRLs in CRCCs, utilizing quasi-static test data from 290 corroded column specimens. In the first stage, a failure mode recognition model and a single-output DRL prediction model were developed using the XGBoost algorithm. This model is then enhanced with the SHAP method to facilitate feature importance ranking and model interpretability. Building on the insights gained from failure mode recognition and feature importance ranking in the first stage, a Deep Neural Network (DNN) was employed in the second stage to achieve multi-output prediction of DRLs. The findings indicate that the SHAP-based interpretable ML method offers profound understanding of the intricate associations between failure modes and DRLs, design parameters and corrosion rate. The proposed DNN model is cap able of concurrently outputting multiple DRLs while balancing the accuracy and efficiency, and signifies a notable advancement beyond traditional methodologies for estimating the lateral deformation capacity of CRCCs. • Established XGBoost single-output models for CRCC failure modes and DRLs. • Conducted feature importance ranking and enhanced model interpretability using SHAP. • Proposed a DNN model for high-accuracy, simultaneous multi-DRL prediction with correct magnitude relationships.