Moderators: Global Moderator, Moderator
Hello,
I am training a force field on various structures, including a surface. According to the best practises page on the wiki, the stresses should not be trained in such a case by setting ML_WTSIF = 1E-10. I forgot to set this flag. Is is possible to make this up in the refitting stage by setting ML_WTSIF = 1E-10 during refitting, or do I have to train the structure again with the right settings?
Hello!
Refitting with a small ML_WTSIF should be fine if you kept the simulation box fixed during your training run. Then the "garbage" stresses did not have a chance to cause any unphysical box changes (which could tear your slab apart, etc.). I would be more careful if you freed the box parameters (e.g. in x,y-directions if the z-axis is perpendicular to the slab). Then it is possible that you visited phase space regions which you could not have reached with small ML_WTSIF. However, even in that case the collected ab initio data can still be correct and you could use it for refitting. I would suggest to visually inspect the training trajectory and look for suspicious behaviour (vast structural changes, defects, etc.). In addition it would be good to check if the convergence for all ab initio steps is good (look for iterations not converging within NELM steps). If all looks ok, I would assume it is safe to use the data for further processing (e.g. refitting).
All the best,
Andreas Singraber