Assessing machine learning prediction of recycled aggregate concrete strength with a tabular foundation model

Recycled aggregate concrete (RAC) supports more sustainable construction, but recycled concrete aggregate (RCA)'s heterogeneity complicates compressive-strength prediction. This study evaluates the Tabular Prior-data Fitted Network (TabPFN) against nine conventional machine-learning models using a standardized literature-derived database of 833 RAC mixtures. Cube strengths were converted to equivalent cylinder strengths, with conversion sensitivity also assessed. Among conventional models, CatBoost achieved the highest accuracy (R² = 0.843), whereas TabPFN achieved R² = 0.870, mean absolute error (MAE) = 3.924 MPa, and root mean squared error (RMSE) = 5.551 MPa. TabPFN also produced a lower RMSE than CatBoost across all four RCA replacement intervals without extensive task-specific hyperparameter tuning. Feature importance and Shapley additive explanation (SHAP) analyses identified effective water-to-cement ratio, aggregate-to-cement ratio, RCA density, and RCA absorption as key predictors, while RCA replacement level alone had limited influence. The moderate improvement over CatBoost indicates that further gains depend on improved RCA characterization.