Kernel-Based Regression in Transient Nonlinear Electro-Quasistatic Field Simulations

During high resolution transient electro-quasistatic field simulations, large sparse nonlinear algebraic systems need to be solved iteratively at each timestep. In previous works, the subspace projection extrapolation method and the Gaussian process regression method succeeded in providing improved start values with known previous transient solutions. In this work, a kernel-based regression model and a linear extrapolation model is combined for providing improved start values for a repeated iterative Newton-Raphson method used within an implicit time- integration scheme. The performance of estimated start values using the combined model on a small data set is presented and compared with other methods.