On 03/10/14 14:15, Gabriel Balaban wrote:
I have personally never seen a case where the solver converges after hitting nan or inf
Hi Gabriel, This is sometimes recoverable with the PETScSNESSolver, so long as the first residual evaluation is not NaN/Inf; if you tell PETSc to use the bt backtracking linesearch, SNES will try backtracking until it finds a stepsize that does yield NaN/Inf. This can allow for SNES to converge where the vanilla Newton solver diverges. For more details and a hyperelasticity example, see https://bitbucket.org/fenics-project/dolfin/pull-request/168/snes-linesearch-fixes/diff If you're getting NaNs in a Newton solve, try SNES with bt. Patrick _______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
