Prediction of creep compliance in high strength concrete using a Feature Tokenizer Transformer with grouped validation

Abstract Long-term creep-compliance prediction is important for serviceability assessment of deformation-sensitive high-strength concrete (HSC) infrastructure, but reported performance can be inflated when related records from the same experimental source are split across training and testing. This study evaluates a Feature Tokenizer Transformer (FT-Transformer) as a material-level surrogate under a leakage-aware, file-grouped benchmark and defines a documented feature interface for future structural-health-monitoring (SHM)/digital-twin-assisted serviceability workflows. A traceable HSC subset was reconstructed from the Northwestern University Infrastructure Technology Institute (NU-ITI) creep and shrinkage database, containing 7606 observations from 432 file groups. Sixteen predictors covering mixture proportions, material type, admixture dosage, strength, geometry, sustained stress, environment, loading history, initial humidity, and log-time transforms were used with common preprocessing and inverse-transformed evaluation. The fixed single FT-Transformer achieved R 2 = 0.9181, MAE = 6.0141 × 10 −6 /MPa, and RMSE = 9.0608 × 10 −6 /MPa on the grouped test set, giving the highest R 2 and lowest RMSE among learned models, while XGBoost produced the lowest MAE. ACI 209 and fib Model Code 2010 served as code-based engineering reference models. Surrogate-assisted SHapley additive explanations (SHAP) identified loading age, loading duration, aggregate-to-cement ratio, and humidity-related descriptors as influential. These results support grouped-database serviceability scenario screening as a material-level analytics component; practical field use still requires project-specific feature mapping, stress estimation, calibration, uncertainty treatment, and structural-response validation.