Machine learning-based collapse prediction for post-earthquake damaged RC columns under subsequent earthquakes.

The rapid assessment for numerous post-earthquake damaged reinforced concrete (RC) structures is critical for making informed post-earthquake decisions. However, this often requires significant resources in terms of manpower and calculation costs. This study proposes a machine learning (ML)-based method for predicting the collapse state of post-earthquake damaged RC columns under subsequent earthquakes, which are a critical earthquake-resistant component in RC structures. First, the damage strain field theory is introduced to establish the numerical analysis models for 10000 post-earthquake damaged RC columns. A dataset is constructed using time history analysis results obtained from these numerical analysis models under the action of 1000 ground motion records. The Random Forest (RF) algorithm is then utilized to train an ML model that can predict the collapse state of post-earthquake damaged RC columns under subsequent earthquakes. To evaluate the accuracy of the ML prediction model, a testing dataset consisting of 47 full-size RC columns from the Pacific Earthquake Engineering Research Center (PEER) database is constructed. Results demonstrate that the prediction model has good accuracy, with an Area Under Curve (AUC) value and an F1 measure of 0.87 and 0.76, respectively. The SHapley Additive exPlanation (SHAP) algorithm is then utilized to interpret the ML prediction model, and the influences of design parameters, ground motion intensity parameters, and damage index on the prediction results are discussed. Finally, an interactive and user-friendly graphical user interface (GUI) tool is developed to provide a rapid collapse prediction of post-earthquake damaged RC columns. This study represents a pioneering step towards the application of ML in post-earthquake assessment for damaged RC columns. • A machine learning-based method for evaluating the collapse safety of post-earthquake damaged RC columns under subsequent earthquakes was developed. • The damage strain field theory is introduced to establish the numerical model of post-earthquake damaged RC column. And the hysteretic behavior of post-earthquake damaged RC column is simulated quantitatively. • A machine learning dataset is constructed using time history analysis results obtained from 10000 post-earthquake damaged RC column numerical models under the action of 1000 ground motion records. • The SHapley Additive exPlanation (SHAP) algorithm is utilized to interpret the machine learning prediction model, and the influences of design parameters, ground motion intensity parameters, and damage index on the prediction results are discussed. [ABSTRACT FROM AUTHOR]