Interpretable Machine Learning for One-Part Fly-Ash/Slag Geopolymer Strength Prediction: Toward Multifunctional...

Area of Science:

• Materials Science
• Civil Engineering
• Chemical Engineering

Background:

• Portland cement production contributes significantly to CO2 emissions (approx. 8%).
• Low-carbon geopolymer binders offer a sustainable alternative.
• One-part (just-add-water) geopolymers enhance safety and logistics for field applications but face formulation challenges due to vast combinatorial possibilities.

Purpose of the Study:

• To review one-part geopolymer science.
• To conduct a comparative and interpretable machine learning (ML) analysis of a pooled dataset of one-part fly ash/ground granulated blast-furnace slag (GGBS) geopolymer mixtures.
• To propose an AI-assisted design framework for geopolymer binders.

Main Methods:

• A literature-pooled dataset of 80 one-part geopolymer mixtures from twelve studies was analyzed.
• Gradient-boosted trees were employed for ML analysis, targeting 28-day compressive strength.
• Leave-one-source-out (LOSO) cross-validation was used to assess model performance on literature-pooled data.

Main Results:

• Gradient-boosted trees achieved an R² of 0.61 (RMSE = 15.5 MPa) under LOSO cross-validation, significantly outperforming a linear baseline (R² = 0.36).
• The precursor balance (fly ash vs. slag content) and activator's Na2O dosage were identified as dominant predictors of compressive strength.
• Non-linear relationships were observed, suggesting transferable structure across different studies.

Conclusions:

• The proposed ML framework demonstrates potential for efficient navigation of the geopolymer formulation space.
• The AI-assisted design framework can guide the development of low-carbon binders, with precursor composition and activator dosage being key factors.
• This framework offers a transferable route toward multifunctional low-carbon binders for various applications, pending experimental validation.