Mixture replication in benchmark datasets inflates reported machine learning performance and degrades prediction interval coverage for ultra-high-performance concrete

Abstract Machine learning models for ultra-high-performance concrete (UHPC) commonly report coefficients of determination above 0.94 on held-out records, and these values now shape expectations for data-driven mix design. We show that a substantial part of this reported predictive performance depends on how the models are evaluated. Benchmark UHPC datasets contain replicate records, meaning the same mixture tested at several curing ages or specimen sizes. Random train-test splitting distributes these replicates across both sides of the split, so models are scored partly on mixtures encountered during training. Using an open dataset of 1,193 UHPC records and an independent benchmark of 1,030 conventional concrete records, we compared random cross-validation with mixture-grouped cross-validation for eight regressors and four target properties. Under mixture-grouped evaluation the coefficient of determination for compressive strength fell from 0.947 to 0.893 and for flexural strength from 0.897 to 0.779, while root-mean-square error rose by 42 to 47% relative to the value implied by random splitting. A controlled experiment in which sample size was held constant and only the number of records per mixture varied reproduced the effect on both datasets, with the gap growing from approximately zero at one record per mixture to 0.13 at four. Conformal prediction intervals calibrated on randomly selected records covered new mixtures 82 to 84% of the time against a 90% nominal level, whereas calibration on held-out mixtures restored approximately nominal coverage with wider intervals. We provide an evaluation and reporting protocol for concrete machine learning.