Machine Learning Model for Determining the 28-day Limit of Concrete Heat Release in Massive Monolithic Foundation Slabs

Introduction. T hermal cracking is a significant challenge for massive monolithic structures. The primary approach to mitigating this risk is to control the overall heat release during cement hydration. Published scientific papers address the direct problem of assessing thermal stresses under given conditions, including the exothermic properties of concrete mixtures. The objective of the study is to solve the inverse problem, that is, to create a reliable machine learning model to predict the 28-day maximum allowable total heat release ( Q max ). Implementation of the proposed approach will enable monitoring the crack resistance of massive foundation slabs. Materials and Methods. To build the machine learning model, a training dataset of 9375 samples was created using finite element parametric modeling. The model takes into account six input parameters: foundation slab thickness, upper surface heat transfer coefficient, ambient temperature, concrete compressive strength class, initial temperature, and concrete hardening rate. To improve forecast accuracy, the CatBoost (categorical boosting) algorithm was used with hyperparameters tuned by Bayesian optimization. Forecast quality was assessed using the mean absolute error, mean absolute percentage error, mean square error, root mean square error, and the coefficient of determination. The value of the parameters for forecasting was determined by a method based on the Shapley theory. Results. The trained model demonstrated high accuracy. The mean absolute percentage error (MAPE) on the training set was 0.52%. The coefficient of determination was 0.99 on the training dataset, and 0.98 — on the independent test dataset. This indicated good agreement within the considered ranges of input parameters. Shap (Shapley additive explanations) analysis identified two key factors for Q max : heat transfer coefficient and slab thickness (with significance values of 0.57 and 0.48, respectively). The parameters ranked by significance were as follows: surface heat transfer — 96%, slab thickness — 75%, ambient temperature — 61%, hardening rate — 35%, concrete grade — 17%, and initial temperature of the mix — 14%. Discussion . The feature correlation was dictated by both the physical processes involved and the modeling methodology. The relatively high correlation between ambient temperature and initial mix temperature (ρ = 0.63) is explained by the conditions under which the model was trained. During dataset construction, the maximum initial concrete temperature was constrained by the ambient temperature. The weak positive correlation between concrete grade and Q max is explained by the ability of high-strength concrete structures to withstand higher temperature gradients. Shap analysis allowed us to clarify: fast hardening at high temperatures reduced Q max , while slow hardening at low temperatures increased it. Conclusion. The machine learning model developed in this study provides an auxiliary computational framework for optimizing concrete composition selection during the design process. Bayesian optimization of hyperparameters provided high quality of algorithm performance in terms of metrics: MAE = 0.67; MSE = 1.78; RMSE = 1.33; MAPE = 0.52%. Three limitations of the model are unaccounted for fast-onset concrete creep, external deformation limitations, and cases of using slabs for transfer structures.