Abstract: As a complex multibody system, artillery serves as the core firepower support equipment of the army. The contact/impact phenomenon arising from the clearance between the barrel and projectile, directly impacts the firing accuracy of the artillery and holds significant research importance. Traditional knowledge-driven modeling methods struggle to balance the accuracy and efficiency when handling the complex surface contact/impact process between the barrel and projectile. This work proposes the integration of the artificial neural network technique with the finite element simulation to develop a neural network-based contact force model. A comprehensive performance evaluation of the established model used for the barrel-projectile contact/impact process will be conducted. Initially, a finite element elastic-plastic dynamics model will be established to analyze the interaction process between the barrel and projectile. Changes of physical quantities under different initial contacting velocities will be analyzed and employed to build a multi-condition sample set for subsequent network training process. After that, a range of evaluation indicators will be introduced to facilitate a thorough and unified assessment of the contact force model, including mean square error (MSE), relative square error (RSE), relative absolute error (RAE), symmetric mean absolute percentage error (sMAPE) and determination coefficient (R²). Considering the significant effects of network hyperparameter settings on the prediction accuracy and stability of the model, the genetic algorithm (GA) will be utilized to obtain optimal hyperparameter settings, focusing on the influence of relevant GA parameters on the prediction performance of the established contact force model. Simulation results demonstrate that with the improved hyperparameter optimization strategy, the prediction accuracy and stability of the proposed neural-network-based contact force model for the interaction process between the barrel and projectile are significantly improved, which further verifies the great necessity of research efforts towards refining hyperparameter settings to enhance the overall performance and generalization ability of the established model.