Machine learning-based hydration heat and relative humidity modeling for optimized design of low-autogenous shrinkage concrete

Experimental testing of autogenous shrinkage, hydration heat, and internal relative humidity in concrete is costly and time-consuming. This study develops machine learning models to predict cumulative hydration heat and internal relative humidity and to analyze the influence of key factors. A dataset of 6300 data points was constructed through interpolation of experimental measurements. Seven machine learning algorithms were evaluated, with the final models based on a fully connected neural network for hydration heat and a long short-term memory network for internal relative humidity, achieving validation biases below 3.5 % and 2.5 %, respectively. Feature importance analysis indicated that water-to-binder ratio primarily affected hydration heat, whereas capillary water content dominated internal relative humidity. Furthermore, a multi-objective mix proportion design framework using the NSGA-II algorithm was proposed to optimize low-autogenous shrinkage concrete by simultaneously satisfying constraints on hydration heat and internal relative humidity. Comparison with existing autogenous shrinkage data confirmed the predictive accuracy of the models, demonstrating their potential for efficient and optimized design of low-autogenous shrinkage concrete.