Preliminary results for “AlkEtOH minus OH set”
OptGeo and Ab Initio targets (i.e. force matching), no VibFreq targets, only Minima
Objective function includes Number degrees of freedom, Number of parameters
slide 4: trace of splitting/merging types and number of types. In the beginning two splitting of types, then merging. Splitting seems to be necessary to go over the obj function hill and then minimize properly to a deeper minimum.
CBy: Can you devide N_{dof} and N_{parm} by temperature and can get something like simulated annealing?
TG: Possibly.
CBy: Plot on Slide 4 looks like Pareto optimization. Also is the splitting due to the penalty function?
TG: No, penalty function will come in later.
CBy: Looks like a multiple minima problem. Due to the two splittings in the beginning and one merging
TG: Landscape is rugged
DM: We have multiple local minima, not all have gradient zero necessarily.
DM: We currently adding and deleting types to find global minima. Could we do that more efficiently?
CBy: We could try simulated annealing to get our of the local minima. We must adopt efficient tactics to jump over barriers.
DM: Instead of doing ForceBalance, do just some parameter moves.
TG: Im doing that already.
DM: You could do some random jumps, following MonteCarlo schemes in Type space. We don’t want to do lots of MonteCarlo steps.
TG: Currently we do 200 to 300 (trial?) splits to find the best splits. Takes about 23 hrs. For each candidate I do a fit and pick the best one. See my tier 1,2,3 scheme
DM: Currently, we do about 500 refits to find split. Could we just have modest amount of Monte Carlo and avoid some of the refits with FB? Imagine you take original torsion and propose a set of starting points and just take 2 or 3 of the best ones and optimize those. That could end-up being better than adding and deleting torsions.
TG: Currently I am also lookgin at bonds and angles.
DM: If you just look at the local minima problem, MonteCarlo is probably the most efficient way to go.
CBy: What is the Delta?
TG: QM vs. FF optimized Geo
CBy: At the state where you evaluate the Objective fuction you got Deltas for whole training set. And then you go all possible changes you could make and the deltas going with them. Could you say a split you want to do drives bigger deltas?
TG: Do all the splits and look what the best drop is. Then I use the best 3 and do a full fit.
CBy: Can you evaluate all splits without going one step? Look at all deltas for all parameters in the force field. Now look at all deltas with all possible splits. Basically just look at the qm data.
TG: This data should already be contained in all gradients from FB.
DM: Chris is talking about splitting. I was thinking about not splitting. Could you improve starting points in order to avoid split/unsplit? Imagine, take each of the parameters, put them in a grid. You would learn that for a given starting point, multiple solutions work. That would be relatively cheep process compared to doing full optimization.
TG: Should we scan all parameters?
DM: Not sure. But if you did, you could find different solutions from the beginning. Now that you know you have a multiple minima problem. You should ask separately if you could do something about the multiple minima problem that is cheaper?
TG: “Cheaper” is the issue.
DM: Initial conditions are not good, don’t have k values.
CBy: The whole drop we see on slide 4, is just normal minimization of parameters and not so much optimization of chemistry
DM: Imagine we have a machine that can optimize typing and one that optimizes parameters and both are expensive to run. We don’t want to run those all the time. Maybe try to change your routine a bit so that you don’t have to run both all the time.
CBy: The remarkable thing you discovered is a huge local minima problem. Do all our FF efforts suffer from this local minima issue.
DM: No, they all have better starting point. However, there is no principle reason why they should not have any local minima problem.
CBy: So this should be an immediate issue to look at.
DM: Potentially.
TG: Bonds were splitted, but were not having effect on bonds/angles. Just on torsions.
CBy: Strange since this is also happening in ring. Is this only for the 1,4 ?
TG: There is 40% force constant difference. If you seperate this out, it won’t improve the force that much.
TG (correction after meeting): r4 bond length split increased from 1.533 to 1.553, with no change in force constant. QM bond length is 1.556
CBy: Trying to understand what happens mechanistically. Seems fishy a bit. In the 4-membered ring (cyclobutane) seems to be under strain. Is it real?
CBn: Did the force constant increase or decrease?
CBy: In a torsion, if you move 1,4 further apart, you would first push out the valence angle.
TG: I can figure out more details on this (see above) and inform you later about this. Next time I should have finished result and show.
0 Comments