Generative Machine Learning Models for Geopolymer Concrete: Strength Prediction and Mixture Optimization

Sabyr Kaiyrbayev, Amanali Bekbolat, Bakbergen Temirzakuly, Essam Shehab, Md. Hazrat Ali

2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.1016/j.procs.2026.06.566
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Geopolymer concrete (GPC) for extrusion-based 3D printing still relies on slow, trial-and-error mix designs to meet target strength and printability. We develop a machine-learning–guided inverse-design framework that couples an XGBoost surrogate with Differential Evolution (DE) under domain constraints (eight mix constituents, box bounds, and a penalized water-to-solid window enforcing printability). Trained on 1,083 labeled mixes, the surrogate attains R² = 0.91 and MSE = 19.56, then drives DE to propose feasible recipes. A benchmarking study shows that the approach outperforms common baselines (SVR, Random Forest, ANN) and aligns favorably with prior literature. A comparative summary is provided in Table 3 to situate our performance and dataset scale among recent studies. Experimental trials of three generated formulations (GEN1–GEN3) demonstrate progressive gains in extrudability, buildability, and 28-day compressive strength, with GEN3 achieving a compressive strength of 31 MPa. The remaining gap between surrogate prediction (≈38 MPa) and measurement underscores sensitivity to dataset composition and curing/printing conditions and motivates dataset curation specific to 3D-printing regimes (rheology-controlled, ambient-cured mixes). Technically, the framework (i) identifies a feasible extrusion window (W/S ≈ 0.33–0.37) that balances flow and early shape retention, (ii) quantifies strength sensitivity to liquid phase and binder partitioning, and (iii) reduces the search space for lab validation while providing interpretable feature effects. We conclude that ML-assisted inverse design can materially accelerate GPC mixture development for additive manufacturing; near-term improvements include ensemble surrogates and uncertainty-aware optimization to tighten prediction–experiment agreement and harden reliability for deployment.