Machine learning-driven predictive models for compressive strength of steel fiber reinforced concrete subjected to high temperatures

Abstract

Steel-fiber-reinforced concrete (SFRC) has emerged as a viable and efficient substitute for traditional concrete in the construction industry. By incorporating steel fibers into the concrete mixture, SFRC offers enhanced crack resistance, improved post-cracking performance, and effective stress transfer. To optimize cost and time in the construction sector, the application of machine learning (ML) methods is now prevalent for accurately estimating concrete characteristics. Accordingly, the present study focuses on utilizing novel ML techniques that include adaptive neuro-fuzzy inference systems (ANFIS), artificial neural networks (ANN), and gene expression programming (GEP) to predict the compressive strength (CS) of SFRC at elevated temperatures. For developing the ML-based models, 307 experimental records were acquired from the published studies conducted between 2000 and 2022. The models' accuracy was assessed using multiple statistical indices. The developed models provide excellent performance with a correlation coefficient (R) value of 0.962 for ANN, 0.998 for GEP, and 0.968 for ANFIS models. Overall, the GEP model provided higher accuracy and less error compared to ANN and ANFIS models. Moreover, the SHapley Additive exPlanation (SHAP) technique was used to interpret the outcomes of the ML-based model predictions. The combined SHAP value of cement content, temperature, and water-to-cement ratio accounts for 80.7% of the total SHAP value across all features. In addition, the SHAP analysis revealed that the actual temperature holds greater significance compared to the heating rate when considering its impact on compressive strength. The comparison of the developed model with the multivariable regression (MLR) method provided that the ML-based models have more prediction accuracy than traditional prediction methods. The proposed models provide designers and builders with an efficient and versatile tool for evaluating attributes, enabling accurate predictions of the CS of SFRC under high temperatures in construction applications.