To address the significant discrepancies between theoretical formulas and experimental results for axial piston pump efficiency under different working conditions, a machine learning-based efficiency calculation method is proposed. First, a nonlinear regression model for axial piston pump efficiency is established, and its validity is verified by significance testing. Subsequently, a predictive model based on a BP neural network is designed, trained and verified using experimental data. Finally, the prediction accuracies of both models are evaluated. The results show that, compared with the existing theoretical formulas under conditions of variable pressure, speed, and flow rate, both the nonlinear regression model and the BP neural network model significantly improve the prediction accuracy. Specifically, the average relative errors is reduced from 8.89% to 1.4% and 0.62%, respectively.