convergence problems in structure optimization

[elena.mol]

Dear all,
I have found convergence problems in geometry optimization runs with both Abinit7.6.2 and Abinit7.10.2, when running in parallel, for several different systems (isolated molecule; adsorbed molecules on a silicon slab with several different cells, molecule type, and adsorption configurations), on the CINECA cluster Galileo (http://www.hpc.cineca.it/hardware/galileo).

After a number of “correct” geometry optimization steps with total energy and "max grad (force/stress)" decreasing (usually 6 to 8 steps for the case I am attaching: molecules adsorbed on a silicon slab, 4x2 cell), the self-consistency starts failing to converge in a reasonable number of steps; then, in the following geom. opt. iterations, the "max grad (force/stress)" and the total energy increase, and after few steps the geometry of my system “explodes” (bonds break etc).

If performing a geom. opt. run starting from the last “correct” geom. opt. iteration of a previous one as defined above, also the new calculation will run with the correct behaviour for 6 to 8 iterations (i.e. beyond the “critical iteration” where the previous one failed), before starting to show the same problems as mentioned above. This procedure can be iterated, yielding reasonable behaviour of total energies and of system geometries. The system is non-metallic, and anyway the same problem is found both with or without using metallic occupations (occopt 3) and a smearing.

No problems were found, instead, on self-consistent calculations on given (reasonable) geometries of the same system, neither in band structure calculations.

In those cases (small gas phase molecule) where it was feasible to perform a serial run on the same machine, or a parallel run on our local linux cluster, the geom. opt. calculations converged regularly. Also, I never encountered these problems with “older” versions of Abinit, which I used on different CINECA machines.


Is the default self-consistency algorithm different for simple SCF runs w.r.t. geometry optimization runs? Can this problem be related to the choice of parallelization keywords, or to the compiling options?

I am attaching the following files (adsorbed molecules on silicon slab):
* gopt28_2THY-Si4x2_2oxbr.in: example geometry optimization input run
* Epot_go28_allsteps: total energies of all SCF steps of each geom opt iteration, from the .log file (i.e. I grepped ETOT)
* max_grad_go28: “max grad (force/stress)” during geom opt, from the .log file
* TEST_THY-Si_U_st10go28sm.xsf.out: coordinates of the system in .xsf format, at the geom opt iteration where the problem appears (see Epot... files)
* config.log: configuration log file for the Abinit version I am using

I am not attaching the geom opt .log file itself, due to its quite large size (10.5 Mb after gzipping), but, if needed, I might extract parts of it.


Thanks a lot in advance
Best regards

Elena Molteni
Department of Physics
Universita' degli Studi di Milano
via Celoria, 16
I-20133, Milano, Italy
and European Theoretical Spectroscopy Facility (ETSF)
http://www.etsf.eu