I have been running a test for Al,Ti,Ni with mmaps using single energy DFT as input.
It all appears to work fine apart from the fact that the list of predicted ground states is really long.
When I plot the convex hull using mmapsrep there are a large set of predicted ground states that has energies far from the facets of the convex hull, see attached figure. Are there a known problem with the ground state identification in mmaps?
I am running the current beta version of the code.
I restarted the calculation to make sure I had not mixed some settings.
It has not run very long but already if I plot the convex hull I find bot predicted and confirmed states below the convex hull.
The problem is that Ti has a composition region that is not fcc, so you may get a lot of structures that relax to something very different, which confuses the fit.
check if some structure have relaxed too much, with checkrelax (>0.08 is suspicious). Do a touch n/error for the offending structures (n).
create a suitable crange.in file that contains concentration limits to exclude non-fcc regions.
See other similar post/responses about this overrelaxation topic.
This is just a test run to validate our setup.
I have done single energy DFT calculations, that is, no geometry optimization at all.
My problems are:
-Not all DFT verified energies are inside the convex hull.
-I get lots of predicted gs that are inside the convex hull.
-I have an example of a convex hull structure with a positive formation energy.
In my opinion there is something wrong with the convex hull routines when run for multi component alloys.
What is the name of the convex hull algorithm? I am trying to read my way through the chull.cc code but its is slow progress since I don’t know what algorithm is implemented.
After some testing the trouble is found to be related to the zero_tolerance setting used.
I changed the chull.c++ code to used chull_zero_tolerance instead of zero_tolerance and set it to 1e-10. The convex hull now looks much better.
If the tolerance is set to high the convex hull routine appear to fail to do proper comparisons for the faces you even end up with faces inside the convex hull.
1e-10 was a too low. It worked fine for a long time then suddenly one of the pure structures where not included in the convex hull. I lowered the zero tolerance to 1e-8 and it now appears to work fine.