Machine learning approaches for lateral strength estimation in squat shear walls: A comparative study and practical implications

Abstract

This study investigated the influence of input parameters on the shear strength of RC squat walls using machine learning (ML) models and finite element method (FEM) analysis. The analyses were conducted on the largest currently available dataset of 639 squat RC walls with a height-to-length ratio of less than or equal to 2.0. The findings suggest that ensemble learning models, specifically XGBoost, CatBoost, GBRT, and RF, are effective in predicting the shear strength of RC short shear walls and using Bayesian Optimization for hyperparameter tuning improves their performance. The axial load had a greater influence on the shear strength than reinforcement ratio, and longitudinal reinforcement had a more significant impact compared to horizontal and vertical reinforcement. The performance of XGBoost model significantly outperforms traditional design models such as ACI 318-19, ASCE/SEI 43-05, and Wood 1990. Additionally, reducing the number of input features from 13 to 10, 8, or 6 still yields reliable predictions with high accuracy. The finding suggests that the use of XGBoost models provides not only comparable accuracy to FEM simulations with non-linear pushover analysis but also the first one can predict the lateral strength in the case of incomplete data which could not be done by FEM. A web application incorporating XGBoost model with various input features can provide valuable insights for predicting the lateral strength of squat shear walls in building structures.