Strength and durability prediction of sustainable concrete incorporating biomedical waste ash and fly ash

There has been an increased push in the research literature pushing the need for sustainable, high-performing concrete. Studies on partial cement replacement with industrial by-products, such as fly ash (FA) and biomedical waste ash (BMWA), have gained traction. However, the models traditionally used to estimate the compressive strength and durability of such concrete systems lag behind in the ability to capture intricate multi-scale interactions when associated with hybrid binder systems. Most of these predictive models are empirical or purely data-driven by neglecting or failing to rely on physical principles and generalization with varying curing conditions and material synergies. This would be addressed by this study with a fully derived ensemble proposing a new framework that incorporates the physics informed and data-driven technique for the prediction and optimization for M25-grade concrete mixes containing FA and BMWA in compressive strength (at 7, 14, 28 and 90 days) and durability prediction. The main psychic module called by this predictive framework is PINN-CSM (Physics Informed Neural Network for Concrete Strength Modeling); it puts physical laws such as Abram’s water-cement relationship and pozzolanic bounds into the loss function of the neural network, improving extrapolative performance and interpretability, achieving an R² ~ 0.96–0.98. Optimization is managed by MODE-STR (Multi-Objective Design Engine for Strength-Tradeoffs and Replacement) using NSGA II for optimal combinations of binders that balance strength, cost, and carbon impact sets. To further strengthen constituent interaction modeling, GNN-CS (Graph Neural Network for Constituent Synergy) leverages a graph-based message passing architecture to capture cross-material synergies while SHAP-FS (Shapley Aware Feature Synthesis) builds interpretable hybrid features with high joint influence on strength sets. It does so by aligning pre-trained models with experimental domains through transfer learning and hydration kinetics ontologies, TLO-CS (Transfer Learning with Ontological Fine-Tuning). This paper also used Random Forest (RF) and Gene Expression Programming (GEP) for evaluating the results under different scenarios. Thus, this integrated framework provides better accuracy and empowers the informed mix design towards low-carbon durable concrete optimized over multiple performance metrics.