Ordinary Portland cement has high energy and CO₂ footprint, motivating the use of geopolymer concrete as a lower-carbon structural binder. However, geopolymer mix design still relies on slow trial-and-error that does not achieve target compressive and flexural strengths. Prior studies used machine learning to predict geopolymer strength, but mostly with few input variables, compressive-strength-only targets and limited algorithm comparison. Hence, there is no clear evidence on which model best predicts strengths for extended geopolymer mix designs. For this study, a database of 497 geopolymer mixes with fourteen mix and curing variables was compiled from literature, and four supervised models’ artificial neural network (ANN), support vector machine (SVM), Gaussian process regression (GPR) and classification and regression tree (CART) were trained on normalized data with tuned hyper-parameters. Model performance was evaluated using correlation coefficient, root mean square error and mean absolute error for separate compressive and flexural targets. All four models achieved high predictive accuracy on test data, with ANN giving the best compressive-strength performance. For flexural strength, SVM provided the most accurate predictions. Different mechanical responses of geopolymer concrete are best represented by different model structures and that multi-parameter machine-learning frameworks can markedly reduce empirical trial batches. Overall, developed comparative machine-learning framework directly addresses limitations of trial-and-error geopolymer mix design by providing practical tool to predict compressive and flexural strengths from mix parameters and to support performance-based, low-carbon mix design. Future work should expand the database to other mechanical and durability properties and embed tools in decision-support software.