Accelerated DGTD Method Employing Hybrid Time Integration Scheme of Runge-Kutta Time Stepping and Neural Network Prediction

In addressing the computational inefficiency arising from complex boundary conditions and multiscale features in large-scale electromagnetic analyses, an accelerated memory efficient discontinuous Galerkin time-domain (DGTD) method utilizing the hybrid time integration scheme based on Runge-Kutta (RK) time stepping and radial basis function neural network (RBFNN) prediction is proposed in this article. First, a dynamic adaptive time step selection scheme is designed to obtain local results, which are then used as the training dataset for global result prediction. Subsequently, the RBFNN method accurately predicts the values of future time steps, significantly reducing the number of time steps required in the computation, thereby substantially improving computational efficiency while maintaining simulation accuracy. In particular, this method also reduces the memory resources required for electromagnetic simulations by capturing the geometric periodicity attributes within the simulation model. Numerical experiments validate the effectiveness and superiority of the proposed method, offering an efficient and accurate computational scheme with embedded neural network prediction for practical electromagnetic problems.