PG-TabM: Domain-Informed Parameter-Efficient Ensemble Learning With Soft Physical Regularization for Concrete Compressive Strength Prediction
IEEE Access · 2026 · 14 · 163677-163704
DOI: 10.1109/access.2026.3744880 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Yize Dai und 1 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Accurate concrete compressive-strength prediction remains difficult when available tabular data are small, compositionally heterogeneous, and unevenly distributed across curing ages and binder systems. Recent studies have already established strong tree ensembles, optimized ensemble learners, creator-variable methods, explainable models, and physics-informed or physics-guided predictors; the remaining problem is therefore not the absence of another nonlinear regressor, but how to introduce defensible material structure without overstating physical fidelity and how to select a compact ensemble whose members provide accurate residual complementarity. PG-TabM addresses this problem as a domain-informed extension of TabM with soft physical regularization rather than as a governing-equation solver. The framework combines material-ratio descriptors, semantic group encoders, age-conditioned residual amplification, support-weighted age-consistency regularization, and a shared parameter-efficient ensemble. Budget- and diversity-aware Bayesian optimization (BD-BO) selects configurations using validation error, cross-fold stability, resource cost, and accuracy-gated residual complementarity. Ensemble-consensus Shapley additive explanations (EC-SHAP) retain the ordinary ensemble-mean attribution while adding member-level dispersion and directional-consensus diagnostics; these diagnostics are treated as explanation-instability indicators rather than calibrated predictive uncertainty. Across three public datasets and ten composition-aware repeated outer holdouts, PG-TabM improves the original TabM, reaching a root-mean-square error (RMSE) of 4.34 MPa, a mean absolute error (MAE) of 3.01 MPa, and a mean absolute percentage error (MAPE) of 8.19% on CCStrength and an RMSE of 8.18 MPa, an MAE of 5.86 MPa, and $R^{2}=0.945$ on UHPC-810. The strongest advantages occur under reduced training coverage, binder-regime shift, and controlled input corruption. The advantage is conditional rather than universal: TabR attains a slightly lower full-data CCStrength RMSE, and CatBoost remains stronger on the 103-sample fixed-age dataset. The repeated outer holdouts are not independent laboratory replications, and no prospective specimen-level external validation is claimed. The evidence therefore supports PG-TabM as a compact domain-informed alternative for small-to-medium, variable-age, compositionally heterogeneous concrete datasets, with external laboratory validation remaining necessary before deployment-level claims can be made.
Zitieren
APA
Dai, Y., & Lu, R. (2026). PG-TabM: Domain-Informed Parameter-Efficient Ensemble Learning With Soft Physical Regularization for Concrete Compressive Strength Prediction. IEEE Access, 14, 163677-163704. https://doi.org/10.1109/access.2026.3744880
Vancouver
Dai Y, Lu R. PG-TabM: Domain-Informed Parameter-Efficient Ensemble Learning With Soft Physical Regularization for Concrete Compressive Strength Prediction. IEEE Access. 2026;14:163677-163704. doi:10.1109/access.2026.3744880
BibTeX
@article{Dai2026,
author = {Dai, Yize and Lu, Rongqing},
title = {PG-TabM: Domain-Informed Parameter-Efficient Ensemble Learning With Soft Physical Regularization for Concrete Compressive Strength Prediction},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {163677--163704},
doi = {10.1109/access.2026.3744880}
}
EUVIMED