A review of machine learning methods for geopolymer performance prediction: Linking data-driven models with physical insights

Abstract

Carbon emissions from the construction industry are becoming increasingly significant worldwide. Geopolymers, as green and low-carbon alternatives to Portland cement, offer great potential for sustainable building applications and have therefore attracted broad research interest. The performance of geopolymers is affected by multiple interacting factors, making conventional trial-and-error approaches costly and inefficient. While machine learning (ML) has been widely applied to predict geopolymer properties, purely data-driven models remain limited by black-box behavior, weak generalization, and dependence on data quality. This review systematically examines (i) the applications and dominant algorithms of data-driven ML in geopolymer research, (ii) the underlying reaction mechanisms that inform physically driven strategies, and (iii) the emerging dual-driven frameworks that integrate data with physics. The dual-driven approach represents a novel paradigm that bridges predictive power with mechanistic understanding, offering both accuracy and interpretability. Evidence shows that ensemble learning and deep neural networks achieve the best stability and robustness in data-driven prediction. Incorporating physical insights enhances interpretability and reliability, while hybrid strategies combining both approaches further improve system stability and accuracy. Unlike previous reviews centered mainly on data-driven models, this paper emphasizes the integration of physical reaction mechanisms with ML and highlights knowledge graph–guided pathways as promising directions. Such integration enables closed-loop optimization through knowledge-guided learning, experimental validation, and feedback correction, ultimately supporting the development of a knowledge–data collaboration platform for geopolymers and advancing sustainable material design.