Multi-objective optimisation of sustainable concrete incorporating solar module waste using an integrated machine learning and NSGA-II framework

The accelerating decommissioning of photovoltaic panels is generating an amorphous-silica-rich waste stream with demonstrated pozzolanic potential in concrete, but conflicting experimental outcomes across replacement strategies demand a computational optimisation framework that does not yet exist. This study develops the first machine learning-driven multi-objective optimisation framework for solar module waste (SMW) concrete, integrating an XGBoost surrogate model with SHAP-based mechanistic interpretation and tri-objective evolutionary optimisation via NSGA-II and NSGA-III. The surrogate was trained on a multi-study dataset compiled from independent experimental programmes and embedded as the fitness evaluator to simultaneously maximize compressive strength, maximize total SMW incorporation, and minimize cement content. SHAP analysis revealed that cement replacement with milled PV glass exerts a net positive effect on strength through pozzolanic C–S–H precipitation, whereas sand replacement degrades performance via disrupted aggregate packing, confirming a bifurcated optimisation landscape unique to this waste stream. The resulting Pareto front demonstrates that cement reductions up to 45% are achievable without breaching 60 MPa when replacement is confined to the pozzolanic pathway, while volumetric waste uptake beyond 1000 kg/m³ incurs unavoidable strength penalties governed by packing physics. Leave-one-study-out cross-validation identified waste characterization standardization as the critical enabler for inter-laboratory model transferability.