Start Value Estimation Using Gaussian Process Regression for Transient Nonlinear Electro-Quasistatic Field Simulations

For high-dimensional transient electro-quasi-static field simulations, large sparse nonlinear algebraic systems of equations need to be solved iteratively at each time step using a solver, such as a preconditioned conjugate gradient (PCG) method. If an improved start value is available, the solver can converge faster, and thus, the simulation time can be reduced. In this article, Gaussian process regression (GPR) is used in order to predict start values for next time steps using data that are collected from previous time steps. Numerical results show that GPR can efficiently predict solutions with good accuracy.