Metaheuristic-optimized machine learning models for predicting compressive strength and assessing sustainability of waste glass powder additive mortars

The importance of waste management, waste recycling and ecological efficiency is felt more deeply every day. In this process where technological advances are developing rapidly, good studies can be carried out with the perspective of multidisciplinary approaches. For this purpose, in this study, it is aimed to evaluate and develop machine learning models optimized with meta-heuristic algorithms for the prediction of compressive strengths of mortars containing waste glass powder (WGP) with high accuracy in terms of sustainability. In this direction, Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CB), Extremely Randomized Trees (ERT) and Random Forest (RF) algorithms were applied using a data set of 281 test result data compiled from 17 different scientific studies; the hyperparameters of these models were optimized with Particle Swarm Optimization (PSO) and Dwarf Mongoose Optimization (DMO) methods. At the same time, the most decisive variables within the mixture parameters were determined according to the SHAP-based variable importance analysis method. The experimental validation study, conducted to assess the real-world applicability and generalization capability of the optimized PSO-RF model, yielded an R 2 value of 0.841 based on comparisons performed on 60 test samples. In addition, the model was supported with sustainability analyses in terms of embodied energy (MJ/kg) and carbon emission (kg CO 2 /kg); the ecological efficiency potential of waste glass powder usage was revealed. According to the obtained data, it was determined that the developed model has high prediction not only in theoretical data sets but also in the application field. • A hybrid ensemble ML–metaheuristic framework predicted WGP mortar strength. • SHAP improved transparency by addressing the black-box problem of ML models • Sustainability of WGP-modified mortars was evaluated via E-Energy and CO 2 metrics • A GUI-based app was developed for strength prediction and sustainability evaluation • Model accuracy and applicability were confirmed through experimental validation