Prediction of Thermal Stresses During the Construction of Massive Monolithic Foundation Slabs Using Temperature Monitoring Data

Thermal cracking during the early-age hardening of massive monolithic foundation slabs poses significant risks to structural durability and integrity. Traditional analytical methods for assessing thermal stresses rely on simplifying assumptions, such as a parabolic temperature distribution, while numerical methods like the finite element method (FEM) are computationally intensive and sensitive to uncertain boundary conditions.

This study proposes a novel methodology integrating real-world temperature monitoring data with machine learning to predict thermal stresses at three characteristic points (bottom, middle, and top) across the slab thickness. A comprehensive dataset of 717,360 records was collected using FEM-based numerical experiments, considering variations in concrete class, slab thickness, curing time, heat release kinetics, and ambient conditions. The CatBoost gradient boosting algorithm was selected for its robustness to multicollinearity and its ability to model complex, nonlinear relationships.

The developed model demonstrated exceptional predictive accuracy, with a coefficient of determination (R²) exceeding 0.99 for all stress components and mean absolute errors (MAE) of 0.130 MPa, 0.383 MPa, and 0.060 MPa for the bottom, middle, and top surfaces, respectively. Feature importance analysis revealed that the heat release rate (99%) and curing time (96%) predominantly influence bottom surface stresses, while the central temperature (94%) governs stresses at the slab’s mid-depth.

Validation against independent experimental data confirmed the model's high fidelity up to the point of crack formation. The proposed approach implicitly accounts for the time-dependent evolution of concrete's elastic modulus through temperature–time parameters, overcoming key limitations of existing analytical solutions.

The proposed methodology enables rapid, near-real-time assessment of thermal stresses directly from in-situ monitoring data, providing a powerful and computationally efficient tool for early-age crack risk management during the construction of massive foundation slabs.