A transformer-based machine learning model for optimizing the design of cementitious mixtures with mine tailings as supplementary cementitious materials

Abstract

Realizing the full potential of incorporating mine tailings as supplementary cementitious materials (SCMs) to replace ordinary Portland cement (OPC) requires carefully balancing the benefits—such as cost reduction and emissions mitigation—while ensuring the mixtures achieve the required strength. Given the demonstrated effectiveness of combining machine learning (ML) with optimization algorithms in similar multi-objective optimization (MOO) problems, for the first time, this study employed a novel tabular prior data fitted network (TabPFN) model to forecast the uniaxial compressive strength (UCS) of those mix designs. The TabPFN model outperformed traditional boosting ML models, achieving an R² of 0.973 and a low prediction error of 2.115 MPa. Notably, its pre-trained architecture reduced computational time by 1045 s. Building on this, a MOO case study was developed using the TabPFN model to predict UCS as the first objective, alongside separate equations used as objective functions to calculate cost and total emissions. This MOO problem was tackled using the non-dominated sorting genetic algorithm-II (NSGA-II). The optimized mixture designs achieved better balances between strength, cost, and emissions than those obtained through experimental methods, validating the use of this ML-based method for mixture design. Finally, a software tool—GreenMix AI—was developed to provide integrated access to the entire framework, translating advanced research into practical application. In essence, this research supports the reuse of mine tailings as SCMs and provides a practical pathway to developing more economical and sustainable cementitious mixtures.