Filler type barely moves the needle on ultra-high-performance concrete strength, database study finds

Ultra-high-performance concrete is one of the strongest building materials ever engineered, routinely reaching compressive strengths above 150 MPa through low water-to-binder ratios, dense particle packing, silica fume reactions and fiber reinforcement. That performance comes at a price: conventional UHPC mixtures often contain more than 800 kilograms of cement per cubic meter, driving up both material cost and embodied carbon. Replacing some of that binder with mineral powders and waste-derived fillers — silica flour, limestone powder, iron tailings, recycled brick powder, glass sand and ground granite powder among them — has become one of the field’s most active research frontiers. Yet a new open-access analysis of more than two thousand published UHPC mixtures suggests that, at the database level, the choice of filler and how much of it you add matters far less than engineers might hope.

The study, published on 28 October 2026, compiled 2073 mixture records drawn from 165 publications spanning 27 countries, using Version 2 of the Global Dataset of UHPC Mix Designs with Supplementary Cementitious Materials, Nano Additives, and Sustainable Fillers. The researchers screened out 115 records lacking a reported 28-day compressive strength and retained values ranging from 80 to 298 MPa, with an overall mean of 150.19 MPa and a standard deviation of 36.37 MPa. Every record kept its original source title, publication year, DOI or web address and country of origin, allowing the team to trace each data point back to its study of origin — a detail that would prove crucial to the findings.

The headline result is striking in its modesty. Of the 2073 mixtures, 360 contained a reported filler and 1713 formed a no-filler reference group. Filler-containing mixtures averaged 148.80 MPa against 150.49 MPa for the no-filler group, a mean difference of just −1.69 MPa with a 95% confidence interval from −5.34 to 1.96 MPa. Welch’s test found no significant overall difference, and the standardized effect size, Cohen’s d and Hedges’ g, was a trivial −0.046. When the 15 harmonized filler categories were compared with one another, analysis of variance and the Kruskal-Wallis test both produced highly significant results, but eta squared was only 0.0295 — filler category explained roughly 3% of the unadjusted strength variation. Statistical significance, in other words, did not translate into a large pooled engineering effect.

Regression analysis told a similar story once major mixture characteristics were held constant. An ordinary least squares model including cement content, silica fume content, water-to-binder ratio, fiber dosage and filler-type indicators explained 35.19% of the strength variation. The pooled filler-dosage coefficient came out at 0.0181 MPa per kilogram per cubic meter, but its heteroscedasticity-consistent p value of 0.0548 missed conventional significance, and the confidence interval spanned zero. A quadratic filler-fraction term added almost nothing, changing explained variance by only 0.0003. The one exception was an exploratory interaction analysis restricted to the 360 filler-containing records: filler-type-by-filler-fraction interactions were jointly significant, suggesting that quantity relationships may differ from material to material even though no universal dosage rule exists.

The most methodologically revealing step was a sensitivity analysis accounting for publication dependence. Because many mixtures in a literature database come from the same laboratory using the same materials and procedures, the records are not independent. When the team re-estimated the same regression with publication-clustered standard errors, several category-specific conclusions changed dramatically. The Silica Flour coefficient, for example, remained 7.00 MPa but its adjusted q value rose from 0.0167 to 0.396, losing significance entirely, while Manufactured Sand also lost significance. Conversely, R-Fa shifted from nonsignificant to significant. Many filler categories were represented by only a single source publication — Balast Powder had 27 records from one publication, R-Fa just ten records from one — meaning that even moderate record counts did not guarantee independent cross-study replication.

Machine learning added a second layer of insight, and a second cautionary tale. Four algorithms — Ridge Regression, Random Forest, CatBoost and XGBoost — were trained on 21 features covering mixture composition, filler descriptors, fiber geometry, curing conditions and testing factors. On a fixed internal holdout of 415 records, XGBoost achieved an R² of 0.8433 and Random Forest 0.8441, figures that look impressive on paper. But the holdout split was not grouped by publication: 414 of the 415 holdout records came from publications already represented during training, so the models were largely interpolating within familiar territory rather than predicting genuinely new studies.

When validation was instead performed with publication-grouped cross-validation, withholding entire source studies, performance collapsed. XGBoost’s mean R² fell to 0.380, CatBoost’s to 0.4115, Ridge’s to 0.3577 and Random Forest’s to 0.2566, with mean absolute errors around 20 to 23 MPa. Nested publication-grouped hyperparameter tuning materially improved Random Forest, lifting its mean R² from 0.273 to 0.438, but changed Ridge, CatBoost and XGBoost only slightly, and tuned grouped performance still remained only about 0.35 to 0.44. The gap between record-level interpolation and transfer to unseen publications is the study’s central methodological warning: high accuracy on a conventional split says little about how a model will fare on mixtures from a laboratory it has never seen.

Explainable AI analysis using SHAP values on the fitted XGBoost model revealed which variables actually drove predictions. Fiber dosage ranked first with a mean absolute SHAP value of 10.148 MPa, followed by curing temperature at 8.251 MPa, total binder content at 5.381 MPa, and specimen size and testing standard at 5.185 and 4.847 MPa. The prominence of specimen geometry and testing procedure reflects heterogeneity in how strength was measured across the source literature. The three filler-related variables ranked far lower: filler type came 16th of 21 features at 0.1928 MPa, filler fraction 18th at 0.1496 MPa, and absolute dosage 19th at 0.1096 MPa. Notably, the broad category descriptor received a larger model attribution than either quantity measure, and this ordering held under a one-hot encoding sensitivity analysis — though the authors stress this is a model-dependent comparison of descriptors, not evidence that filler identity is intrinsically more important than dosage, and SHAP attributions are not statistical significance or causality.

The authors are careful about what these patterns do and do not show. The database lacked consistent measurements of particle-size distribution, specific surface area, mineral composition, morphology, pretreatment and reactivity, so filler type functioned as a broad categorical label rather than a true material characterization. Literature datasets may also carry publication bias, since successful mixtures are reported more often than failed trials. The study examined only 28-day compressive strength, leaving workability, shrinkage, flexural behavior, durability and environmental performance unaddressed. The authors frame their contribution as source-conditioned database evidence for hypothesis generation — identifying which questions deserve controlled, material-specific experiments — rather than universal filler rankings, causal effects or engineering-ready prediction tools.

For a field searching for sustainable ways to cut cement use, the message is double-edged. On one hand, the small pooled filler effect suggests that swapping in mineral or waste-derived fillers need not dramatically compromise 28-day strength, at least across the range of published mixtures. On the other hand, neither the statistical associations nor the machine-learning attributions can tell an engineer which specific filler to choose or at what dosage, because within-category variation is large, independent replication is scarce, and predictive models transfer poorly to unseen research groups. The authors recommend independent validation on newly acquired experimental data, fuller filler characterization and case-specific uncertainty quantification before any engineering deployment — and, in the meantime, treat every apparent filler ranking from a literature database with healthy skepticism.

Subject of Research: Statistical and machine-learning analysis of filler type and dosage in relation to 28-day compressive strength of ultra-high-performance concrete

Article Title: Statistical and machine learning analysis of filler type and dosage in relation to 28-day UHPC compressive strength

Article References: Jiang, X., Yang, J., Chen, X., & Yang, X. (2026). Statistical and machine learning analysis of filler type and dosage in relation to 28-day UHPC compressive strength. Discover Concrete and Cement, 2(1), Article 92. https://doi.org/10.1007/s44416-026-00134-9

Image Credits: AI Generated

DOI: 10.1007/s44416-026-00134-9

Keywords: ultra-high-performance concrete, UHPC, compressive strength, fillers, machine learning, XGBoost, SHAP, publication-grouped cross-validation, supplementary cementitious materials, cement reduction, regression analysis, construction materials