<p>To address the issues of subjectivity and insufficient optimality in the segmentation positions of gradient virtual material in the dynamic modeling of bolted joints, this paper proposes a method for optimizing the segmentation positions. First, based on the distribution characteristics of contact pressure in the bolted joint, the variation range of each segmentation position is determined, and the corresponding analytical model for the gradient virtual material parameters is established. Next, by combining the deep neural network (DNN) surrogate model, a mapping relationship is developed between the gradient virtual material parameters at different segmentation positions and the dynamic characteristics of the joint. Finally, the results of the modal analysis experiments are compared with the predictions of the DNN model using an optimization algorithm for the identification of the gradient virtual material parameters and the optimal segmentation positions. A multi-bolted connection plate structure is used as the research object, with three different gradient virtual material models (two-segment, three-segment, and four-segment) are established, and the optimal segmentation positions are identified. The results show that the accuracy of all three models improves after optimizing the segmentation positions, with the highest accuracy achieved when the model is divided into four segments. In this case, the relative errors of the first six order natural frequencies are less than 0.9951 %.</p>