Machine learning-based prediction of torsional behavior for ultra-high-performance concrete beams with variable cross-sectional shapes

Abstract

Ultra-high-performance concrete (UHPC) is renowned for its exceptional mechanical properties; however, its torsional behavior remains inadequately understood, posing challenges for its application in structures subjected to twisting loads. Existing prediction methods often fall short of accurately capturing the complex interplay between material characteristics, cross-sectional geometry, and reinforcement, leading to significant errors. This work introduces a unique Machine Learning (ML) method to accurately anticipate the torsional behavior of UHPCs. Three powerful algorithms, Random Forest, Gradient Boosting Regressor, and Long Short-Term Memory (LSTM), were trained and assessed on a dataset of 113 UHPC specimens. The best R-squared was 99 % provided by the Gradient Boosting Regressor, while the LSTM and Random Forest showed 98 % and 96 % accuracy. The ML approach determined that splitting tensile strength, fiber length, web width, and stirrup diameter were the most important factors controlling torsional force. These results provide insight into the complex interaction affecting UHPC torsional performance, opening the path for accurate UHPC design in challenging applications.