In-plane lateral strength prediction of unreinforced masonry walls subjected to lateral loads using optimized boosting algorithms

Journal: Structures, WoS(SCIE), Q1
Author: Trương Gia Toại - Trường Đại học Đông Á

Abstract:This study conducts an in-depth evaluation of the effectiveness of machine learning (ML) models optimized using particle swarm optimization (PSO) in predicting the in-plane lateral strength of unreinforced masonry (URM) walls for engineering purposes. A dataset comprising 136 specimens with six key input variables is employed to train the models. Baseline ML algorithms, including gradient boosting (GB), extreme gradient boosting (XGBoost), light gradient-boosting machine (LightGBM), and categorical boosting (CatBoost), are utilized with default hyperparameter settings as provided by the sklearn library. In contrast, hybrid models combining PSO and ML algorithms (PSO-GB, PSO-XGBoost, PSO-LightGBM, and PSO-CatBoost) are developed, leveraging PSO to optimize the hyperparameters of the baseline models. The predictive performance of each model is thoroughly evaluated using various metrics to identify the best-performing model, which is then compared with existing design methodologies. Additionally, the interpretability of the forecasting models is enhanced through Shapley Additive Explanation (SHAP) analysis, addressing the "black box" nature of ML techniques. A parametric study was also performed to exam the influences of variations in input variables on the in-plane lateral strength of URM walls. To further facilitate practical implementation, a graphical user interface (GUI) is developed based on the most accurate forecasting model, enabling straightforward prediction of the in-plane lateral strength of URM walls. The findings reveal that PSO-ML hybrid models significantly enhance prediction accuracy compared to default ML models, with testing R² ranging from 0.926 to 0.954 versus 0.811–0.941. Among them, the PSO-CatBoost model achieved the highest performance (R² = 0.954, RMSE = 24.299 kN, MAE = 16.093 kN, and MAPE = 13.942 %). Furthermore, the PSO-CatBoost model outperformed existing design equations in predicting the in-plane lateral strength of URM walls, yielding a mean experiment-to-prediction ratio of 1.00 and a lower COV of 0.07. SHAP analysis highlights that the wall length (L_w) is the most critical factor influencing the lateral strength, providing valuable insights for engineers in design and evaluation processes. Moreover, the optimal combination identified was the interaction between wall length and wall thickness. These results underscore the potential of PSO as a robust optimization technique for enhancing ML model performance in structural engineering applications.

DOI: In-plane lateral strength prediction of unreinforced masonry walls subjected to lateral loads using optimized boosting algorithms - ScienceDirect