Compressive Strength Prediction for Ultra-High-Performance Concrete Mixture Design Using a Hybrid Role-Aware Tabular Learning Method

Predicting compressive strength from constituent quantities and curing conditions supports the assessment of candidate mixtures in ultra-high-performance concrete (UHPC) mixture design. Existing machine learning models usually process recorded variables directly or require composite variables that depend on field semantics. This study proposes SARN–TabPFN, a hybrid architecture that integrates a schema-agnostic role-aware neural representation (SARN) with version 2 of the Tabular Prior-Data Fitted Network (TabPFN-v2). SARN identifies a repeated experimental structure, initializes latent material roles from numerical distributions, and learns normalized role descriptors through a joint prediction and regularization objective. The hybrid representation combines these descriptors with the standardized variables without requiring complex feature engineering. The method was evaluated on 792 processed records from a public UHPC database against random forest, extreme gradient boosting (XGBoost), and identically standardized TabPFN-v2 using identical partitions. The experiments show that SARN–TabPFN achieves the highest mean R2 and lowest mean RMSE, MAE, and MAPE among the evaluated models under both shuffled ten-fold cross-validation and ten-fold cross-validation in which the test base mixtures are not included in the training set. Under the latter protocol, its mean MAE was 5.804 MPa compared with 7.010 MPa for TabPFN-v2. These results support the potential use of SARN–TabPFN to assess candidate mixtures against a specified compressive strength requirement within the compositions and curing conditions represented in the database.