Machine learning based determination of lateral displacement of fully grouted reinforced shear masonry walls under in-plane axial loading

Abstract

Reinforced shear masonry (RSM) walls are critical for resisting lateral loads, but accurately estimating their displacement capacity is challenging due to single reinforcement layer. This study addresses this issue by developing and comparing three machine learning algorithms: Convolutional Neural Network (CNN), Gene Expression Programming (GEP), and Tree-Structured Parzen Estimator-Extreme Gradient Boosting (TPE-XGB), to predict the lateral displacement of fully grouted RSM walls. A database of 152 experimental instances was compiled, incorporating variables like masonry strength, reinforcement ratios, axial load ratio, and shear stress demand. The TPE-XGB model demonstrated the highest predictive accuracy (R² = 0.995), followed by CNN, while GEP provided a less accurate but interpretable empirical equation (R² = 0.884). Comparisons with simpler regression models confirmed the efficiency of the machine learning approaches. To enhance transparency, Shapley Additive (SHAP) and Individual Conditional Expectation (ICE) analyses were conducted, identifying aspect ratio, shear stress demand, and grouted masonry strength as key factors influencing displacement. An explainable computational tool was also created to facilitate practical implementation, enabling engineers to accurately predict the lateral displacement behavior of RSM walls.