Advanced computational models for accurate fracture toughness prediction in diverse concrete types: Insights from a robust laboratory database

Abstract

Fracture toughness (FT) is one of the most important material characteristics that determines the mechanical stability and operational performance of concrete structures, but its accurate measurement is rather difficult due to the complicated, expensive, and lengthy nature of the common experimental approaches. This study proposes a novel approach that combines machine learning (ML) with empirical information from three-point bending tests to estimate FT for a large set of concrete formulations. A dataset of 600 samples, which were carefully prepared and tested, containing aggregates of different sizes, water-cement ratios, curing conditions, and material additives, was used. Compared to other ML algorithms, the AD-XGBoost model achieved the best results by predicting FT with R² values of 0.99 and NRMSE of 0.0008. As per the sensitivity analyses, cement type, fiber percentage, and curing temperature were the most influential parameters on FT, while the inclusion of carbon and steel fibers into the concrete matrix enhanced its structural stability. Improvement of the toughness of concrete was also noticed with a prolonged curing process. Besides being an effective and reliable substitute for the prediction of the FT, this ML modelling strategy significantly minimizes the need for resource-demanding laboratory tests.