Skip to end of metadata
Go to start of metadata

You are viewing an old version of this page. View the current version.

Compare with Current View Page History

Version 1 Current »

Background

Hyesu evaluated suitable priors for use in fitting for OpenFF 1.3.0: https://openforcefield.atlassian.net/wiki/spaces/FF/pages/723025921/v1.3.0+release+note#4.-Determination-of-priors-width-of-valence-parameters

However, now that we fit to the initial guesses obtained through the MSM method, Pavan has found these priors to be unsuitable. OpenFF 2.1.0 used the following priors:

priors
   Angles/Angle/k :  100.0
   Angles/Angle/angle :  5.0
   Bonds/Bond/k :  100.0
   Bonds/Bond/length :  0.1
   ProperTorsions/Proper/k :  5.0
   ImproperTorsions/Improper/k :  5.0
/priors

These were a bit too loose (pasted from Slack):

Trevor & I were looking at the parameter changes wrt the initial MSM starting point for 2.1.0 and the use of larger priors shifted some force constants by a larger value since the mval changes during a FB run are not penalized as much if the priors are large. Especially, the k value of a32 , the hypervalent sulfur angle, changed by 44% from 241 units to 134 units, which is almost 1*mval change with a prior of 100 units on angle force constant. So, the angle was not as stiff as the initial point. It would be counter intuitive to use larger priors if our intention is to be as close to the QM informed starting point as possible. So I tried two more iterations with tighter priors and the one I picked for 2.2.0 is a good one with no degradation in benchmarks.

He evaluated multiple sets of priors, finding the below to be too tight:

priors
   Angles/Angle/k :  5.0
   Angles/Angle/angle :  1.0
   Bonds/Bond/k :  10.0
   Bonds/Bond/length :  0.01
   ProperTorsions/Proper/k :  5.0
   ImproperTorsions/Improper/k :  5.0
/priors

But the below set gave good results, and was used in : 

priors
   Angles/Angle/k :  20.0
   Angles/Angle/angle :  1.0
   Bonds/Bond/k :  20.0
   Bonds/Bond/length :  0.01
   ProperTorsions/Proper/k :  5.0
   ImproperTorsions/Improper/k :  5.0
/priors

Goal

Repeating Hyesu’s analysis with the new MSM-derived initial values may be worthwhile for determining the most optimal set of priors, or at least for systematic evidence that this set of priors is the best.

  • No labels