Open Access

ARTICLE

Machine Learning for Compressive Strength Prediction of 3D-Printed Concrete: Feature Engineering, Statistically Validated Model Selection, and Applicability Boundaries

Jia Chen¹, Zhicheng Liao¹, Mengdi Hou², Jianbo Huang^1,3,*

1 School of Computer Applications, Guilin University of Technology, Guilin, China
2 Guangxi Key Laboratory of Machine Vision and Intelligent Control, Wuzhou University, Wuzhou, China
3 Centrale Méditerranée, Technopôle de Château-Gombert, 38 rue Frédéric Joliot-Curie, Marseille, France

* Corresponding Author: Jianbo Huang. Email:

(This article belongs to the Special Issue: Frontiers in Computational Modeling and Simulation of Concrete)

Computer Modeling in Engineering & Sciences 2026, 148(2), 12 https://doi.org/10.32604/cmes.2026.085729

Received 16 May 2026; Accepted 21 July 2026; Issue published 28 August 2026

Abstract

Extrusion-based 3D-printed concrete (3DPC) imposes a dual constraint on mix design: fresh-state printability and hardened compressive strength must both be maintained within a narrow water-to-binder window, making data-driven prediction tools essential for reducing experimental iteration. This study evaluates 20 regression algorithms on 254 experimental records spanning plain printable mortars to high-fibre reinforced composites (CS: 11.1–189.0 MPa). Four physically motivated composite variables encoding cement blend potency, cumulative supplementary cementitious material (SCM) substitution, fibre volumetric stiffness, and water-to-sand ratio are constructed; Boruta-based selection retains 11 of 17 candidate features. CatBoost achieves the highest 30-run mean performance ($R^2=0.8968\pm 0.0505$; single-run metrics on the seed-42 partition: RMSE = 8.82 MPa, MAPE = 9.68%) under frequentist Bonferroni correction (${\alpha }^{\ast }=0.0026$) and Bayesian Savage-Dickey comparison; XGBoost is the only alternative not statistically separated from CatBoost after correction ($p=0.085$; ${\mathrm{B}\mathrm{F}}_{10}=9.23$ indicates only moderate evidence for a small advantage). Multi-method interpretability analysis identifies the water-to-binder ratio as the dominant predictor with an ALE effect range of 57.57 MPa; local interpretable model-agnostic explanations reveal a sign reversal of this effect across the strength spectrum, capturing the printability–strength coupling specific to layer-by-layer deposition. For practical mix design, W/B $\le 0.28$ consistently projects CS above 80 MPa within the compiled dataset; the 95% prediction interval half-width of 12 MPa is recommended as a design safety margin. Trained model is well-calibrated for OPC/SAC-based 3DPC within the compiled W/B window (0.15–0.65), achieving MAPE $\le 9.68$%. Cross-system transfer to 262 ECC/SHCC specimens yields a 12.0 MPa systematic overestimation attributable to material-system differences rather than model overfitting, delineating the applicability boundary for deployment. A graphical user interface implementing the optimized model is released to support mix design without programming expertise. Source codes are publicly available at https://github.com/lucassivan/ML-3DPC.

---

Keywords

---