About differences – there might be more reasons: One can be newly created boundary which is dependant on specific mesh. If new boundary is smoother it gives more efficiently used elements whereas if it is bumpy it “handicap” elements. Another, harder to explain simply, is manner of using stress distribution and its filtering. Stress is taken as average from integration points (more for 2nd order elements) and taken as if it acts in the centre of gravity, which decreases advantage of 2nd order elements. Results of filtering in each iteration are given by filter range. Here I used range as 2 times element size, but there might be some more convenient values for different element types.Very interesting that different elements give different results, even if you use hundreds of them. Would be interesting what the reason for this is. Mhh neiter nature nor materials does not know femelements. Which one is best ?! It is may be like a path. If the optimizer is on a specific path it just goes this way. There must be academic papers about this.
Mhh and tria3 give a better result than tria6 and quad4 (IMHO) really strange ...
I should definitely read more in literature.