Predicting Compaction Characteristics of Geopolymer-Stabilised Soils Using Chemistry-Informed Machine Learning

Abstract

Compaction characteristics are critical for the design and field control of geopolymer-stabilised soils. However, most previous studies have focused on strength, with limited attention to optimum moisture content (OMC) and maximum dry density (MDD). This study developed a chemistry-informed machine-learning (ML) framework to predict OMC and MDD using soil plasticity index (PI), binder dosage, and oxide-based binder descriptors. A database of 144 stabilised soil mixtures was compiled from past published studies, covering soils with PI values from 15 to 60. Linear regression (LR), random forest (RF), multilayer perceptron (MLP), CatBoost (CAB), and XGBoost (XGB) models were developed. The nonlinear models performed better than linear regression. XGB was selected for further analysis because it gave the best balance of accuracy, robustness, and interpretability. The results showed that PI and binder dosage are the primary controls on compaction behaviour, while oxide composition provides useful secondary refinement. Increasing plasticity and binder dosage generally increased OMC and reduced MDD. These findings suggest that chemistry-informed ML can support faster mix design and reduce repeated trial-and-error laboratory compaction testing for geopolymer-stabilised soils.