Concrete manufacturing consumes vast quantities of natural resources and contributes significantly to environmental degradation and carbon emissions. Therefore, integrating recycled waste substances into concrete has become a crucial approach to fostering eco-friendly building practices and supporting circular economy concepts. This study investigates the potential of incorporating recycled aluminum beverage can crumbs (RABCC) as a partial replacement for natural coarse aggregates (NCA) in concrete mixtures, focusing on its impact on compressive strength (CS) and the feasibility of its application in structural concrete. A comprehensive experimental program was conducted to assess the mechanical properties of concrete with varying levels of RABCC (up to 30%) and its interaction with other mix parameters, including water-to-cement ratio (w/c), superplasticizer content (SPC), silica fume content (SFC), and curing time (CT), fine aggregate content (FAC), coarse aggregate content (CAC), fly-ash content (FA–C). The experimental findings revealed that low levels of RABCC incorporation (≤5%) led to minimal reductions in CS, with strengths comparable to those of the reference mix. However, at higher replacement levels, significant reductions in CS were observed, with the CS decreasing by up to 29.8% at 20% RABCC. To predict CS across different RABCC contents, a series of machine learning models, including kernel-based methods, ensemble tree models, gradient boosting techniques, and neural networks, were developed and validated using both hold-out and 5-fold cross-validation. The Gaussian process regression (GPR) model demonstrated the best performance, achieving an R² of 0.88 and an root mean squared error (RMSE) of 3.33 MPa in hold-out testing, and an R² of 0.89–0.94 and an RMSE of 2.37–3.10 MPa in 5-fold cross-validation, confirming its robustness in predicting CS. Additionally, SHapley Additive exPlanations (SHAP) analysis identified the w/c ratio and RABCC content (RABCC–C) as the most influential factors on CS, with RABCC–C exhibiting a moderately negative correlation. This study demonstrated the potential of RABCC as an eco-friendly, sustainable alternative to conventional aggregates in concrete, offering a viable pathway to reduce aluminum waste without significantly compromising the material’s mechanical performance. The results also underscored the importance of optimizing RABCC content to balance sustainability goals with structural performance.

Soilcrete is a composite material of soil and cement that is highly valued in the construction industry. Accurate measurement of its mechanical properties is essential, but laboratory testing methods are expensive, time-consuming, and include inaccuracies. Machine learning (ML) algorithms provide a more efficient alternative for this purpose, so after assessment with a statistical extraction method, ML algorithms including back-propagation neural network (BPNN), K-nearest neighbor (KNN), radial basis function (RBF), feed-forward neural networks (FFNN), and support vector regression (SVR) for predicting the uniaxial compressive strength (UCS) of soilcrete, were proposed in this study. The developed models in this study were optimized using an optimization technique, gradient descent (GD), throughout the analysis (direct optimization for neural networks and indirect optimization for other models corresponding to their hyperparameters). After doing laboratory analysis, data pre-preprocessing, and data-processing analysis, a database including 600 soilcrete specimens was gathered, which includes two different soil types (clay and limestone) and metakaolin as a mineral additive. 80% of the database was used for the training set and 20% for testing, considering eight input parameters, including metakaolin content, soil type, superplasticizer content, water-to-binder ratio, shrinkage, binder, density, and ultrasonic velocity. The analysis showed that most algorithms performed well in the prediction, with BPNN, KNN, and RBF having higher accuracy compared to others (R² = 0.95, 0.95, 0.92, respectively). Based on this evaluation, it was observed that all models show an acceptable accuracy rate in prediction (RMSE: BPNN = 0.11, FFNN = 0.24, KNN = 0.05, SVR = 0.06, RBF = 0.05, MAD: BPNN = 0.006, FFNN = 0.012, KNN = 0.008, SVR = 0.006, RBF = 0.009). The ML importance ranking-sensitivity analysis indicated that all input parameters influence the UCS of soilcrete, especially the water-to-binder ratio and density, which have the most impact.