PB: Last week we talked about whether mixing cyclic and acyclic molecules is going to affect adversely on the parameters, it is not so in the case of linear and cyclo alkanes. (here) Coming to sulfonamide case, larger gradients are observed for these molecules with 1.3.0 and there are no cases of large gradients with the opt-only-targets-refit.
CB: I am not expecting the sulfonamide away from the O-,OH to be distorted so much.
SB: I think a couple of points we can glean from this, one is that there is something wrong with the fitting targets, where there are strong electrostatics puckering the molecules, and one other thing is we should look at the sulfonamide set since in the opt geo targets we include lot of conformers' minimized geometries, it would be interesting to see if these over abundance is driving the optimization to nonsensical places.
TG: Yeah, there is uneven distribution of conformers.
CB: Yeah, but how we ended up here is still puzzling, but the QM structures seem reasonable.
SB: Yeah, 1.2 is probably fine and after that during optimization we sort of passed a barrier and landed in this zone.
SB: Yeah, my guess is it can be a combination of things like optgeo, vibfreq driving forcebalance in a bad direction.
CB: Could there be a chemistry space where this is doing pretty well and that’s why it optimized to that value?
PB: Actually, with a simple fit of linear alkanes the optimizer goes out of the bounds of values seen in the training set, that is because of a bad initial conditions as in here it went to a local minima.
CB: Ohh, so acyclic alkanes HCH goes to 90 degrees, which should be a well defined chemical space.
DLM: Yeah, that’s concerning.
TG: Having the obj. fn values you can say that it is a local minima.
CB: I am still puzzled on what is going well that it stuck here in this local minima, if we can understand this then we can make progress.
SB: Yeah, I agree. PB, you have the parameters from these fits right, so take those and minimize using openmm and compare with QM strcutures once again. FB sees RMSDs in internal coordinates and comes up with optimized parameters, so if we don’t see any outliers then there is something else that’s going kaput.
PB: Okay.
SB: Also, we should double check the internal coordinate RMSDs in FB, it can also be a minor software bug.
DLM: Yeah, exactly, we should check whether the angle RMSDs are going down or not too.
CB: PB, we now have a small chemical space and we can reproduce the pathology, so we can now sort this out.
PB: One last thing to show is that the torsion drives from the td-only-refit, where there is a drop in force constant, and 1.3.0 almost overlap with each other except in one case.
TG: Probably, the huge bend in the angles drags k down.
CB: Are we confident that the chemistries that drive a30, a31 are not adversely impacting the parameters, it might be a global parameter?
TG: I can answer that, a31 has a large spread of angles with chemsitries other than sulfonamides, such as four membered and five membered rings. When I split on these my code splits the sulfonamides completely, taking even the Nitrogen (in place of oxygen) ones, and these pushing and pulling causing bad parameters.
CB: So, when splitting based on equilibrium values endocyclic angles are separated?
TG: So, the equilibrium angles based splitting does split out the endocyclic out but I didn’t fully do it.
CB: So, I think that’s significant, looking back at PB’s geometries where there are O=S=O with higher angles, your conclusions agree with that.
JM: One quick question on the Taft-Hammett constants, which one to pick from the table?
CB: I think the sigma_p is good, the one for the para substituent.
JM: Okay.
DLM: Regarding XTB plots from JM’s undergrad we’re looking at the correlation between AM1-QM, and XTB-QM, we will get the RMSE, etc., sooner or later. So far, XTB looks like it has a better correlation with QM bond order. Our AM1 is dependent on the calculation procedure and software bottlenecks, is this a potential replacement for AM1 in the future, may be XTB-BCC, etc.,?
CB: I would be happy if we find XTB is a better method, then I am more than happy, but we may face the same problem we have with describing conjugation, charges, with any electronic structure method, so we need to be careful I guess.
DLM: Yeah, right now AM1 has a different implementation in different toolkits and it is a challenge to get parity. The problems are not due to inherent model as you said, I agree.
CB: Yeah, getting the chemistry to do the right thing is priority whichever method we choose, we still have to do ELF10 kind of stuff, etc…
Action items
Pavan Behara will compare the QM, MM geometries of the alkanes with bad-parameter-FF.
Pavan Behara Checkout the geometries of the sulfonamide outliers optimizing with 1.2.0.
Pavan Behara Check out FB code for opt geo targets
Add Comment