KMAN Publication Inc. AI and Tech in Behavioral and Social Sciences In Press 2026 10 22 Prediction of the Compressive Strength of Graphene Oxide-Containing Concrete Using Machine Learning Models and a Deep Neural Network Prediction of the Compressive Strength of Graphene Oxide-Containing Concrete Using Machine Learning Models and a Deep Neural Network EN Journal Article <p>The aim of this study was to develop a comparative computational framework for predicting the compressive strength of control and additive-modified concrete mixtures and to compare the performance of six algorithms with different levels of complexity. The dataset comprised 27 actual laboratory observations from three mixture families, C0, CG, and CP, evaluated at ages of 7, 28, and 90 days with three replicates for each condition. To prevent information leakage and obtain a more realistic estimate of model generalizability, each mixture type-age combination was treated as an independent group, and the models were evaluated using three-fold group cross-validation. In each fold, 18 actual observations were allocated for training and 9 for testing. Data augmentation was applied exclusively to the training set, and five synthetic samples were generated for each actual observation using controlled Gaussian perturbation. After removing variables with no variability, three features-sample age, graphene oxide content, and the recorded polymer-modifier content-were retained as effective modeling variables. Out-of-fold prediction results showed that support vector regression achieved the best performance among the evaluated models, with a coefficient of determination of 0.771, a root mean squared error of 2.059 MPa, and a mean absolute error of 1.641 MPa, followed by gradient boosting and linear regression. The Friedman test, based on the mean errors of the nine independent groups, confirmed a statistically significant overall difference among the models (χ² = 27.159, p &lt; 0.001). In addition, the Wilcoxon test with Holm correction showed that the superiority of support vector regression over the deep neural network, Extra Trees, and Random Forest was statistically significant; however, its differences from linear regression and gradient boosting were not significant after multiple-comparison correction. Permutation importance analysis for the support vector regression model further showed that the recorded polymer-modifier content, sample age, and graphene oxide content accounted for approximately 57.0%, 30.1%, and 12.9%, respectively, of the performance degradation caused by feature permutation. Overall, the findings indicate that, in small datasets, proper validation design, prevention of information leakage, and group-based evaluation play a more decisive role in obtaining reliable predictions than merely increasing model complexity.</p>