Abstract: A method for predicting the assembly accuracy of multi-stage rotors, which takes into account both geometric errors and non-uniform loads, was proposed. First, the assembly deformation caused by the non-uniform load was superimposed on the classical model which only considers position and direction deviations to establish a more accurate assembly error propagation model of multi-stage rotors. Then, the finite element method was used to simulate rotor assembly accuracy under non-uniform loads. The dataset was constructed by FEM, and the KAN neural network surrogate model was trained to accurately predict key parameters in the error propagation model. The proposed surrogate model was compared with the traditional regression model, and its superiority and reliability in accuracy prediction were verified. Finally, a set of random error parameters was used as a case study. The prediction values of the proposed model were compared with those of the classical model. The preload of the joint surface was optimized by the firefly algorithm, and the coaxiality was reduced by 38.03% compared with that before optimization, which significantly improved the assembly prediction accuracy of the rotor system.