Information-based model reduction for nonlinear electro-quasistatic problems

Abstract

We suggest a nonlinear model reduction approach for transient electro-quasistatic field simulations of high-voltage devices that comprise strongly nonlinear electric field stress grading material. The singular value decomposition is employed to obtain the proper orthogonal decomposition modes, while nodes at which interpolation constraints are imposed, are sampled according to a greedy approach and to information criteria. More precisely, in addition to the greedy approach, at each node of the computational mesh the spectral Shannon entropy of the electric potential is computed and interpolation constraints at high-entropy nodes are introduced. Numerical experiments validate that this maximal information node sampling strategy results in improved reduced models, in terms of nonlinear iterations and, in cases, also in terms of accuracy.