This paper introduces a novel three-dimensional electromagnetic (EM) solver, grounded in theory and enhanced by integrating a convolutional neural network (CNN) with the finite-difference time-domain (FDTD) method. The proposed solver is designed for efficient and precise full-wave simulation of large-scale structures. By substituting the traditional operators in the FDTD method with convolutional operators, our approach maintains the accuracy and stability inherent to the FDTD method, while also being ideally suited for parallel computations. Compared to existing models, our proposed CNN- FDTD solver demonstrates improved accuracy, efficiency, and flexibility. Numerical validations confirm its superior stability and computational efficiency.