Prediction of flexural crack width of reinforced concrete beams using interpretable machine learning algorithms

Accurate crack prediction is crucial for structural systems' serviceability limit state design, yet it remains a challenging task. This study explores the use of machine learning (ML) algorithms to predict the flexural crack width in reinforced concrete (RC) beams with an extensive data set of 675 experimental samples. Each training sample consists of five predictive features: strength of concrete (fck), stress in steel (fs), diameter of the bar(dbt), ratio of reinforcement (p ), cover depth (c ), and the corresponding target feature (maximum crack width) uwcr. Analysis and comparison of six different ML models reveal that the Extra Gradient Boosting Regressor, with a mean error of 18 percent, was the best performer among the ML algorithms considered in this study to predict the crack width. Additionally, SHapley Additive explanations have been used to explain the trends observed in the ML models and the identification of significant parameters. It is noted that the crack width is significantly influenced by the stress in reinforcing steel. To assess the predictions using the standard international codes of practice, five codes have been considered: fib MC 2010, Eurocode (EC) 2, Proposed EC2, ACI, and IS 3370. It is seen that the EC2, proposed EC2, ACI code and fib MC 2010 are able to predict the crack width with reasonable accuracy, having a mean error of 53 percent, and the proposed ML Models can enhance the crack width prediction in RC beams.