Nonlinear field simulation with finite element domain decomposition methods on massively parallel computers (Q676377)

The numerical solution of nonlinear boundary value problems on massively parallel computers with MIMD architecture is discussed. The nonlinearity is treated by Newton's method combined with the nested iteration principle (``full multilevel method''). This approach allows to keep the number of Newton iterations small on the fine grids having many unknowns. The coarse grid solution which can be computed with a low numerical effort serves as a good initial approximation for the iteration on the finer grids. For solving the systems of linear algebraic equations within each Newton step, the preconditioned conjugate gradient (CG) method is applied. Here, a domain decomposition (DD) preconditioner is used [see, e.g., \textit{G. Haase, U. Langer} and \textit{A. Meyer}, Computing 47, No. 2, 137-151, 153-167, (1991; Zbl 0741.65091, Zbl 0741.65092)]. Furthermore, the coupled finite element/boundary element discretization of linear and nonlinear problems is discussed. For solving the corresponding systems of linear equations within each Newton step, the CG method of \textit{J. H. Bramble} and \textit{J. E. Pasciak} [Math. Comput. 50, No. 181, 1-17 (1988; Zbl 0643.65017)] is used. The construction of a finite element/boundary element DD preconditioner is described. Finally, the efficiency of the proposed methods on parallel computers with MIMD architecture is shown by three numerical examples, namely two magnetic field problems and an optimal design problem.