Switch TP VLE solver to Powell's dogleg method. Might be just a little bit more reliable, but hard to tell.
