Compressive-Strength Prediction and Generalization of Solid-Waste Alkali-Activated Materials Under Multi-Source Data Heterogeneity
Compressive strength in solid-waste-based alkali-activated materials (AAMs) depends on precursor chemistry, activator composition, curing history, age, and specimen geometry, producing substantial heterogeneity in literature-compiled data. This study examines how validation boundaries affect apparent predictive accuracy and reliability. We analyzed 1367 records representing 907 unique mixtures across nine database-defined subdomains. We compared hierarchical feature sets and assessed grouped cross-validation, leave-one-domain-out (LODO), mixture-purged LODO, complete age/geometry hold-outs, within-stratum transfer, and zero-shot external testing. Adding processing and chemistry information raised XGBoost grouped-validation R2 from 0.518 to 0.843. Across five tree-based models, LODO increased RMSE by an average factor of 1.86; purging repeated mixtures reduced pooled R2 from 0.454 to 0.274. In 70 external observations from 12 literature-source groups, zero-shot R2 was −0.289 and RMSE was 16.20 MPa. Across five training–calibration repetitions, nominal 90% split-conformal intervals attained mean external coverage of only 36.3%, while 48.6% exceeded at least one training-feature bound. These findings show why predictive accuracy, input support, and uncertainty coverage must be evaluated together before transferring AAM strength models across data domains.
Authors
- Rui He (ORCID: https://orcid.org/0000-0001-9088-0774)
- Xuanxi Li
- Wei Chen
- Yinbo Zhang
- Xingyang Gao
Institutions
- Chang'an University (CN)
- Xinjiang University (CN)
Publication Details
- Journal
- Materials
- Published
- 2026-11-19
- DOI
- https://doi.org/10.3390/ma19224894
- Primary Topic
- Concrete and Cement Materials Research
- Type
- article
- Field-Weighted Citation Impact
- 0.00