Data-driven surrogate models with uncertainty quantification for lightweight foamed concrete

This study introduces a data-intensive framework designed for the prediction of compressive strength and dry density of lightweight foamed concrete (LFC) by deploying five advanced machine learning (ML) models: CatBoost, NGBoost, PySR, TabNet, and XGBoost. A specially assembled database of 191 different mix designs was employed, and the models' performance was assessed through rigorous statistical metrics (R2, MAE, RMSE, MPAR) combined with k-fold cross-validation for robustness purposes. Among the tested models, XGBoost achieved the highest overall predictive accuracy (R2=0.994 and RMSE=26.09 kg/m3 for dry density), while NGBoost offered nearly comparable performance for compressive strength (R2=0.983, RMSE=1.82 MPa) and uniquely provided predictive distributions enabling uncertainty-aware design and reliability analysis. In order to maximize explainability, SHAP (Shapley Additive Explanations) analysis revealed cement content and foam volume as the top drivers of strength and density, verifying compliance with known engineering rules of thumb. Symbolic regression (PySR) yielded interpretable equations that approximate the structure-property relationships. As an application example, NGBoost was embedded into a minimalistic graphical user interface (GUI), and engineers can simply feed mix parameters for probabilistic predictions instantly. The proposed framework combines accuracy, interpretability, and usability and highlights the ability of ML surrogates for the acceleration of experimental mechanics and support for stochastic simulation, reliability analysis, and auto-mix design. Although limited by the relatively small and heterogeneous dataset, this research advances the computational simulation of sustainable concretes and complements ongoing efforts for the integration of the field of structural materials engineering with artificial intelligence.