Ensemble machine learning-based approach with genetic algorithm optimization for predicting bond strength and failure mode in concrete-GFRP mat anchorage interface

Abstract

Glass fiber-reinforced polymer (GFRP) bar reinforced concrete structures are susceptible to bonding failure because of the low bond strength between GFRP bars and concrete. In this study, four tree-based machine learning models have been used to predict the flexural bond strength and failure mode of mat anchorage between concrete and sand-coated GFRP bars. Machine learning models are Decision Tree, Random Forest, AdaBoost, and XGBoost; except for Decision Tree, the models were inspired by collective learning. After applying these models to the dataset, the R2 score of the test scores for AdaBoost, Random Forest, XGBoost, and Decision Tree models were 0.91, 0.88, 0.90 and 0.88, respectively. Then, to improve the performance of these models, genetic algorithm was used to optimize the hyperparameters, XGBoost, Decision Tree, Random Forest, and AdaBoost which led to an increase in R2 score by 2, 3, 4, and 3 percent, respectively. Also, XGBoost classification model was used to predict the failure mode, and all the test data (70%) were correctly predicted. In the end, the SHapley value technique was used to determine each feature’s effect on the best machine learning model for predicting adhesion stress and the classifier model for predicting failure mode.