Explainable AI-based optimization framework for sustainable fly ash concrete design

Fly ash is widely recognized as an effective cement replacement material for reducing the environmental impact of concrete production while maintaining acceptable mechanical performance. This study develops an explainable artificial intelligence-based multi-objective framework for the sustainable design of fly ash concrete by combining predictive modeling, feature interpretation, uncertainty assessment, and optimization within a single workflow. A compiled dataset of 1,062 concrete mix samples was used, from which outlier detection was applied only to the training subset, resulting in 58 anomalous samples being removed from the training data while the independent test set remained unchanged. Four artificial intelligence models, namely TabNet, SVR, NGBoost, and SAINT, were trained and compared for predicting 28-day compressive strength. Among them, NGBoost achieved the best testing performance, with RMSE = 4.897, MAE = 3.492, MAPE = 7.844, and R² = 0.92. Feature interpretation showed that cement content, water content, and MCS-28 were the most influential variables governing strength prediction. To further evaluate model reliability, a bootstrap-based uncertainty analysis was conducted, showing that TabNet and NGBoost exhibited comparatively narrower uncertainty bounds. The best-performing predictive model was then coupled with four nature-inspired optimization algorithms to generate concrete mix designs that maximize compressive strength while minimizing CO₂ footprint. The optimization results demonstrated that the proposed framework identifies theoretical Pareto-optimal candidate mixtures within the dataset domain, highlighting its potential as a data-driven decision-support tool for sustainable concrete mix design. The optimized mixtures are computational Pareto-optimal solutions and not experimentally validated production mixes. A limitation of this study is the lack of source-level labels in the compiled dataset, which prevented grouped validation across original studies. Future research may incorporate additional practical constraints, such as water-to-binder ratio and workability requirements, and extend the framework to larger and more diverse datasets.