Work

Domain-aware machine learning for concrete compressive strength prediction with calibrated uncertainty

Year: 2026
Type: article
Abstract: Machine learning is promising for predicting concrete compressive strength, but global point accuracy does not establish reliability for every mixture. This study evaluates a workflow combining regression, prediction intervals and an applicability-domain warning using the public University of California, Irvine (UCI) dataset. Removing 25 duplicate records retained 1,005 observations. The historical 804/201 development/holdout partition was retained for auditing; model and feature selection were repeated exclusively within 643 proper-training observations, leaving 161 observations for calibration. Repeated nested validation and grouped mixture validation supplemented the historical holdout. Gradient boosting with engineering-derived variables obtained a holdout RMSE of 4.16 MPa after fitting all development observations. The separately trained conformal predictor achieved 92.5% empirical coverage at 90% nominal coverage, with an untruncated symmetric interval width of 14.59 MPa; the implemented nonnegative-support intervals had a mean width of 14.57 MPa. Coverage was 78.6% at least 60 MPa and 63.6% outside the distance domain. The paired bootstrap RMSE interval for raw-feature gradient boosting versus raw-feature Extra Trees included zero; supplementary Table S4 reports paired development-validation MAE results. Domain-error differences varied across random splits. These findings support joint reporting of accuracy, interval diagnostics and feature-space support, while limiting claims to internal benchmarking and requiring independent production validation.
Language: English
Open Access status: hybrid
Estimated APC paid: $3,090
OpenAlex ID: W7227322993
FWCI:
Cites: 19
Cited by:
Related to: 10