Multi-target machine learning-assisted design of sustainable steel fibre-reinforced concrete

This paper addresses noted limitations in the machine learning (ML)-based steel fibre-reinforced concrete (SFRC) design. Existing studies having a single-targeted approach have overlooked the primary purpose of fibres of enhancing the flexural strength. Although this effect has been considered in the few available multi-target studies, the role of data structures in data-driven modelling was not taken into consideration. These gaps have led to material designs that can show significant deviations from the targeted sustainability and mechanical goals of multi-objective optimisation (MOO). We propose an ensemble ML trained on a comprehensive dataset of 502 mixes with 15 independent variables to simultaneously estimate the mechanical properties of SFRC. After investigating the accuracy of the developed model and the physical integrity of the learned input-output relationships with SHapley Additive exPlanations, the model was deployed into a mixed variable MOO framework aware of common statistical issues to generate optimum mixes. The framework can capture the interdependence among variables of the training data, known as multicollinearity, and constrain the search space, avoiding low certainty estimated data points. The proposed framework enabled the identification of optimal mixes with a reduced global warming potential by up to similar to 30 % and marginal changes in the cost of production compared to conventional mixes across the modulus of rupture classes of 6, 8, and 10 MPa. The performance of the framework showed an error of less than 6.2 % against advanced numerical simulations and further emphasises the potential for optimal material identification in industry applications.