Concrete, produced from cement, is the best greatly utilised building material. However, greenhouse gas discharges from cement preparation and consumption cause significant damage to the environment. Geopolymer production, which is one of the important alternatives, plays an important role in preventing this problem. In this study, tree-based machine learning (ML) algorithms such as Gradient Boosting Regression (GBR), Decision Tree (DT), Extremely Randomized Tree (ET), and Random Forest (RF) were utilized to anticipate the compressive strength (CS) of silica fume substituted obsidian-based two-component geopolymer mortars with different alkali activator properties. These ML algorithms were implemented using different train-test ratios (0.6 − 0.4, 0.7 − 0.3, 0.8 − 0.2, 0.9 − 0.1). The prediction and generalization performances of the applied models were measured by applying different statistical metrics like R2, MAE, MAPE, MSE and RMSE. For the prediction of compressive strength, the GBR algorithm showed a better prediction performance than the other algorithms, with an R2 value of 0.972. The RF algorithm showed the most consistent and balanced prediction performance. Significant decreases in R2adjusted values were observed as the training rate increased. This is due to the tendency of the models to overlearn as the training rate increases. The results show that the models perform best at a training rate of 70%, and the generalization execution of the models reduces importantly as the training rate augments. The machine learning method applied to the forecasting of the CS of geopolymer mortars provides significant benefits to engineering applications due to its contributions in terms of workload and time savings.