Comparative analysis of machine learning algorithms for predicting shear strength of URM walls

Abstract

Unreinforced masonry (URM) walls are commonly used as primary load-bearing elements in seismic regions, making the assessment of their shear strengths crucial for design considerations. The traditional method for determining shear strength via experimental studies is both costly and time-consuming. This study aims to predict in-plane shear strength of URM walls using Machine Learning (ML) algorithms. For this purpose, a total of 93 tested URM specimens were collected and evaluated, with the results analyzed through Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost) and Gene Expression Programming (GEP) algorithms. The results reveal that the GBM algorithm was the most accurate, with R2 value of 0.931. Additionally, a closed-form model based on GEP to predict the shear strength of URM walls, with an R2 value of 0.916 is also introduced. SHapley Additive exPlanations (SHAP) analysis is conducted on the GBM model to demonstrate the contribution of each input parameter. The SHAP analysis indicates that the length of masonry wall (Lw) has the most significant impact on the shear strength of the URM wall, while its compressive strength (fm) has the least. The findings of this study are expected to enable the application of ML algorithms in predicting the shear strength capacities of URM walls.