ABINIT Discussion Forums

Dear abinit users and developers,
2administrator: Sorry, I didn't found the proper section for my question, so I decided to ask here, please move the topic in the proper section of abinit forum.
Here is my prehistory: I'm using abinit half of year or less and found some convergence study problems I'd like to figure out. When I perform structure optimization I set the tolmxf(The main parameter under which structural relaxation iterations will stop) to value of 1.0d-10. Usually abinit calculations fails to converge. Then I rise tolmxf and start calculation again. Some times calculations succeeds with tolmxf=1.d-8 but usually convergence succeeds with tolmxf=1.d-6 only (I even have structures with structure calculations converges only with tolmxf=1.d-3 ).
And here is my question. What is the optimal value of tolmxf? Why some structures converged with tolmxf=1.d-8 and some only with tolmxf=1.d-3?
Many thx in advance!!!

Hello,

The input variable tolmxf does not refer to a convergence. It is related to the relaxation of your structure. When you are trying to optimize a cell to its equilibrium volume and geometry (which would mean a force equal to 0 on every atom and a zero stress tensor), in fact you give a value below which you consider it is "enough close" to 0. This is the tolmxf parameter : when all the forces on all the atoms are less than the value of tolmxf, then no further move of the atoms or of the unit cell is applied. For the stress tensor, its a bit different, its not anymore values with Newton units (force) that must be close to 0 but it is values with Pressure units (GPa or else) that must be close to 0. So internally, the criterium used to stop the optimization of the cell uses tolmxf*strfact with strfact being another input variable (the default value, 100 is fine).

Concerning the convergence of the SCF cycle itself within each molecular move is something else. For this you can use several criterium to stop the SCF cycles : toldfe, tolvrs, tolwfr, toldff, tolrff. From these ones, toldff and tolrff are maybe the most adapted ones for structural relaxation (except if your system has a symmetry such that the forces on each atom is always 0 !!! so be careful and in this case use another one with a small value, e.g. tolvrs 1.0d-12 or less).

I hope it helps you

David

Dear David, many thx for your detailed answer. When I talked about convergence with tolmxf I did not mean SCF cycles stop criteria. I was actually meant of tolmxf as a stop criteria at equilibrium volume and geometry search. And the problem I report is what some systems are able to reach the state with minimal force at values 10.0d-8 while other are not. Well, am I wrong thinking that the minimal values of forces for each system differ, and differ a lot?

Hello again,

In principle, there is a configuration for which all the forces are 0. In practice, for large systems, the number of degrees of freedom is so large that it is very difficult to get close to it. Usually a value like the default one (5d-5) is already very good. If you cannot go beyond 1d-3 try to continue the relaxation. One trick is also to first relax the structure without relaxing the unit cell (optcell 0 and ionmov /= 0) and then relax both starting with the coordinates you get after having relaxed the positions.

David

Nice trick David, Thank you!!!