Skip to Content
BuildingsBuildings
  • Article
  • Open Access

19 November 2026

23 Pages

Interpretable Machine Learning for Coupled Strength–Flowability Analysis and Screening of UHPC Based on Orthogonal Experiments

,
,
,
and
1
School of Mine Safety, University of Emergency Management, Sanhe 065201, China
2
School of Future Cities, University of Science and Technology Beijing, Beijing 100083, China
3
Department of Civil and Environmental Engineering, George Mason University, Fairfax, VA 22030, USA
*
Author to whom correspondence should be addressed.

Abstract

An orthogonal experimental design was used to systematically investigate the combined effects of four UHPC mix-design factors—silica fume content, fly ash content, quartz powder content, and sand-to-binder ratio—on compressive strength at 3 d following steam curing and 28-day standard curing, as well as on fresh-state flowability. A total of 900 mixture-proportion records that passed the absolute-volume closure check were compiled, of which 556 fell within the adopted workability range of 200–250 mm. A dual-response machine-learning framework incorporating physical feasibility screening was developed; six regression models were compared, and SHAP was used for model interpretation. In the orthogonal experiments, silica fume content exhibited the largest response ranges for both early-age strength and flowability within the investigated factor levels. Mixture A4 achieved the highest measured compressive strengths of 129.76 MPa at 3 d following steam curing and 119.42 MPa after 28-day standard curing, but its flow diameter of 182 mm fell outside the adopted workability window of 200–250 mm. Under this workability constraint, A6 achieved the highest 3-day steam-cured strength among the workable mixtures, and A2 had the highest 28-day standard-cured strength. The results show that the highest-strength tested mixture is not necessarily the mixture with the best strength–flowability balance. LightGBM provided the best prediction for flowability, with a test-set coefficient of determination (R2) of 0.9147 and root mean square error (RMSE) of 7.9203 mm; the gradient boosting model provided the best prediction for 28-day compressive strength, with R2 of 0.8597 and RMSE of 5.9846 MPa. SHAP analysis indicated that the water-to-binder ratio (w/b) exhibited the largest global attribution to predicted strength, with higher values generally associated with lower predicted strength. Higher values of paste volume fraction (Vp), w/b, and superplasticizer-solids-to-binder ratio (SPs/B) were generally associated with increased predicted flowability. These findings highlight the need to consider strength and flowability jointly in UHPC mix design.

Article Metrics

Citations

Article Access Statistics

Article metric data becomes available approximately 24 hours after publication online.