AbstractAs a promising environmentally friendly construction material that can be used to replace concrete, fly ash-based geopolymer (FABG) should meet the working strength requirement. However, the optimal mixture design of FABG could be difficult to obtain through experimental methods due to a variety of influential factors and their complex interrelationships. To address this problem and explore the influence patterns of those factors, this study developed an ensemble machine learning modeling method that integrated three algorithms: support vector regressor (SVR), random forest regressor (RFR) and extreme gradient boosting (XGBoost). A database containing 896 experimental instances was constructed by reviewing open resources. During the modeling, established estimators were tuned with a metaheuristic algorithm called differential evolution (DE). After analysis, the XGBoost model was determined as the strength prediction model of FABG, because it showed the best performance with the largest R2 scores (0.97 and 0.91) without overfitting by the minimum mean absolute error (MAE) gap between the training and testing subsets. Additionally, a further understanding of how the factors affect the predicted values of the model was given by the SHapley Additive exPlanations (SHAP) theory. The results show that curing conditions had the biggest impact on the model output, followed by alkali-activator solution variables and the mole of sodium hydroxide. Therefore, the proposed method can accurately predict the strength of produced FABG and assist in understanding the influence patterns of various factors.