This problem seems to be getting a lot of research lately. The problem being how to process PDEs on complex domains because meshing is a bitch. There seem to be a number of good approaches, NURBS Fea, this, Lawerence Livermore and Caltech just come out with a polygonal mesher.
I wonder how tractable and accurate they are from the solver side. Because generating high quality meshes really is a pain in the ass.