Ensemble machine learning prediction of compressive strength in waste-derived sulfoaluminate cement paste

Industrial solid wastes are increasingly used as alternative feedstocks for synthesising sulfoaluminate cement (SAC). However, their complexity in compositions leads to unstable performance. To optimise production, machine learning (ML) models are developed to predict the compressive strength of SAC pastes based on a dataset of 707 datapoints from literature. Distinct from traditional mineral-based approaches, this model incorporates multi-source factors including feedstock composition, clinker calcination temperature and duration time, gypsum type and content, specimen preparation conditions, and curing time. Single and ensemble ML approaches, including Random Forests (RF), Supporting Vector Regression (SVR), and Neural Network (NN) algorithms, are employed. The ensemble RF + NN model demonstrates higher accuracy (testing R² = 0.87) than the single models. Model-based interpretation reveals that feedstock composition is the foremost input feature group that accounts for 34.9 % importance, thereby validating the composition-driven prediction strategy. Moreover, the correlations of each input feature with compressive strength have been analysed. The ensemble ML model is validated through 14 independent experiments on SAC paste samples prepared exclusively from hazardous waste, with all prediction errors well below 10.82 %. This work provides a precise, data-driven tool for rapid feedstock screening and process optimisation, offering a labour-saving and cost-effective pathway to accelerate sustainable SAC production.