A predictable fractal derivative Maxwell concrete creep model was proposed to address the problems of complex calculations and limited prediction accuracy for multi-source dispersed data in existing concrete creep models. The fractal derivative theory was combined with the XGBoost machine learning algorithm through a physics-data fusion approach. A mapping relationship between concrete material characteristics, environmental conditions, and the key parameters of the fractal derivative Maxwell model (fractal order and viscosity coefficient) was established based on 746 sets of experimental data selected from the NU concrete creep database. Bayesian optimization was adopted to determine the optimal hyperparameters. The constructed model achieved coefficients of determination (R²) of 0.919 and 0.908 for the prediction of the fractal order and the viscosity coefficient on the test set, respectively. An R² score of 0.909 was obtained on an independent validation set. These results demonstrate that the model has good fitting performance and generalization capability. The proposed physics-data fusion model improves overall prediction accuracy while retaining theoretical interpretability compared with existing models.