LBFGS Convergence Criterion

[ucecmtd]

Hello,

Does anybody know if there is a different convergence criterion when using LBFGS as the optimizer in a dimer calculation?

The reason I ask is as follows:

I have a calculation where I run a preliminary calculation with lowered plane wave kinetic energy, SCF tolerance and k-point sampling using the FIRE optimizer to get an initial guess to be used in a more stringent calculation. With this initial run, I ensure that the curvature is consistently negative and the force on the dimer drops until it is low-ish.

Now using the geometry and unit vector along the dimer axis from the preliminary calculation, I increased the cutoff, SCF tolerance (EDIFF = 1e-7) and k sampling (note: these parameters give well-converged energies for stable state calculations) whilst switching to the LBFGS optimizer (with default settings). This new calculation converged with all forces on non-fixed atoms dropping below my desired accuracy of EDIFFG = -0.001. However, the dimer force does not converge to less than this - in fact it appears to be over an order of magnitude greater (see below) despite the calculation stopping due to reaching the convergence criteria...

Step Force Torque Energy Curvature Angle
1 0.04872 26.20388 -89.70996 13.89656 30.02478
1 0.04872 7.50580 -89.70996 -1.21436 16.09547
1 0.04872 6.12038 -89.70996 -3.16422 7.89352
1 0.04872 2.91902 -89.70996 -3.98803 5.69917
2 0.04872 --- -89.70996 --- ---

When using the dimer in the past I have always used the FIRE or QM optimizers and this is the first time I've used LBFGS - FIRE and QM, in my experience always converge to Forces on dimer < -EDIFFG. Does this optimizer have different convergence criteria? Why does the dimer force not converge with the forces on the atoms?

Thanks, any help or suggestions are very much appreciated :)

[graeme]

If you can make the files for these calculations available (both the FIRE calculation and then the subsequent LBFGS calculation), I can then take a closer look at what's going wrong.

Also, just to make sure I understand, you are saying that the dimer force printed (0.04872) is significantly higher than your convergence criterion (0.001) and yet it converged after the first step -- does that sum it up?

[ucecmtd]

Sure, I've attached them to this post - though checking the calculation I used the CG instead of FIRE for the preliminary calculation. The issue with LBFGS is the same though.

Yeah, that's right - the forces on each free-to-move atom in the OUTCAR have converged to be lower than -EDIFFG (0.001), however the dimer force as reported in the DIMCAR is much higher (0.04872).

Thanks for taking a look.

[moore1]

BFGS is a pretty reliable method
https://www.everipedia.com/Limited-memory_BFGS/