Ultra-High-Performance Concrete (UHPC) is increasingly used in slender and prestressed structural members due to its superior strength and durability. However, inaccurate or incomplete prediction of creep deformation may lead to excessive long-term deflection, prestress loss, cracking, and potential serviceability or safety risks in buildings and infrastructure. Therefore, reliable prediction methods for UHPC creep are essential for both structural design and long-term performance assessment. In this study, a database containing 60 literature-derived UHPC creep records was compiled to investigate the creep coefficient at approximately 100 days. Pearson correlation analysis revealed strong interdependence among predictors and weak single-variable linear relationships, indicating that creep behavior is governed by nonlinear interactions. A feedforward backpropagation neural network (BPNN) trained using the Levenberg–Marquardt algorithm was developed to predict the creep coefficient. To maintain engineering interpretability, the fib Model Code 2020 (MC2020) formulation was adopted as a code-based benchmark and further calibrated using ridge regression. Results show that the calibrated MC2020 model improves prediction consistency, while the BPNN model provides the highest predictive accuracy. The proposed framework integrates machine-learning prediction with interpretable code-based calibration, contributing to the development of creep modeling approaches for UHPC and providing practical support for the safe design of UHPC structures.
Early-age autogenous shrinkage remains a major obstacle to the durable use of ultra-high-performance concrete (UHPC), because the response is controlled by coupled mixture, curing, age and restraint effects. This study develops a data-driven prediction workflow from an external database containing 7678 UHPC shrinkage records and nine mixture, mechanical and curing descriptors. Mask-related columns were removed, true zero contents for supplementary cementitious materials and steel fibers were retained, and nonphysical zeros were imputed within each training fold. Pearson analysis showed weak single-variable correlations with shrinkage (|r| < 0.20), confirming the need for nonlinear models. BPNN, RF and XGBoost were evaluated using shuffled five-fold cross-validation and a high-shrinkage-sensitive training workflow. Based on the reported prediction-comparison dataset, RF and XGBoost achieved the strongest global agreement, with R 2 values of 0.823 and RMSE values close to 112 με, whereas BPNN gave R 2 = 0.774, RMSE = 126.62 με and MAE = 63.95 με. The high-shrinkage tail, uniformly defined as measured shrinkage greater than 1,000 με, remained more difficult: for these 226 records, the mean biases were −109.9, −111.6 and −128.6 με for BPNN, RF and XGBoost, respectively. These results indicate that UHPC shrinkage prediction from literature-derived curve points requires both global accuracy reporting and explicit tail-error checking. The model was further linked to bridge-deck monitoring data to distinguish material-level free autogenous shrinkage from restrained field response.