Abstract: Carbon Fiber Reinforced Polymer (CFRP) composites are widely used in aerospace applications due to their high specific strength and specific modulus. However, panel components are prone to develop assembly stresses during the assembly process because of their low rigidity and large dimensions, which may subsequently affect the structural service performance. To address this, this paper proposes an assembly stress prediction method based on a Finite Element Model and Multi-Layer Perceptron (FEM-MLP).The method first uses strain measured by Digital Image Correlation (DIC) as a benchmark to calibrate the constitutive parameters of the finite element model through the Particle Swarm Optimization algorithm. A deep learning mapping network (RegStressNet) is then introduced to train the relationship between DIC strain and FE stress, enabling rapid prediction of FE stress from DIC strain. This method excels in predicting multiple stress features, achieving an average normalized root mean square error of 13.63% for maximum principal stress prediction, aligning with the industry's demand for more precise prediction techniques. Further transfer experiment verification indicates that the model maintains an error level of 16.63% in predicting the dominant stress components of cross-size specimens, demonstrating a certain degree of engineering generalization capability.