Physics-supervised ensemble learning model for predicting failure modes of reinforced concrete columns

Abstract

In order to overcome the limitation that traditional prediction models lack of learning ability of physical laws and with low generalization performance, a novel physics-supervised ensemble learning model (PELM) for predicting failure modes of reinforced concrete (RC) columns was proposed. An ensemble learning model for predicting failure modes of RC columns was established first based on the extremely randomized trees (ET) algorithm. Then the physical laws between feature parameters and different failure modes of RC columns were explored. Meanwhile, a new strategy for hyper-parameters optimization of the PELM was developed based on the physical laws and the K-fold cross validation algorithm of equal proportion. Finally, the learning ability of physical laws and the generalization performance of the proposed PELM were validated by comparing with traditional machine learning (ML) models. Analysis results show that the prediction accuracy of the proposed model can be improved by about 7–18 % and the recall for flexure-shear failure can be improved by about 10–37 % compared with traditional ML models. Moreover, the prediction results of the proposed model are consistent with the physical laws between feature parameters and failure modes, which provides an efficient approach to reflect the flexure-shear competitive relationships for different failure modes of RC columns. Furthermore, the proposed model overcomes the overfitting problem, which relies less on training samples and achieves higher prediction accuracy as well as better generalization performance.