Test fits
PB: Here are some test fits done in the context of sulfonamide debug.
First one is to check the difference in optimized parameter values with a bunch of linear and cyclic alkanes, where the standard deviation of parameter values is low when different subsets of targets are chosen, which kind of implies that cyclic targets won’t move the values far away from the global fit.
Second set of fits are for the sulfonamide where a particular type of targets are chosen to fit among TD, OPT, and VIB targets (listed on the page linked above).
CB: So, form the optimizations it looks like the wild-west parameters in Frostt99 are not a far shot form them.
a30b looks like a meaningful child parameter from the angles being drastically different, although the k-values are not that different.
So, no matter what, around the hypervalent sulfurs we have some angles squeezed and some being large.
So, in your fits splitting a30 resulted in no change in angle but a change in force constant, and the child parameter has a different angle and force constant, that’s interesting.
SB: From table 1 we see that ring structures are not driving the equilibrium angles to lower values, for example cyclobutane mixed with others…
PB: Actually, cyclo -butane and -propane were assigned different angle parameters for C-C-C
SB: Ohk, then we need to check that hypothesis with a simpler forcefield.
PB: Okay, will do that.
SB: Also, these fits are with the harmonic restraints (priors) applied, right?
PB: Yeah, they’re.
SB: Okay, then may be one more refit can be done without the priors would be good.
So, either the ring systems driving the angle smaller, or the optimization, or something orthogonal to both
CB: What do you think about a30b child parameter?
SB: There’s something strange here, the optimization of a single parameter has a lower force constant and the split parameters both have a higher force constant.
CB: PB, can you also do another refit with column 2 having the child parameter?
PB: Sure.
SB: Another striking thing is the low force constant for the TD targets only fit, I wonder how the residuals look like for that, may be there is a subset of torsiondrives that are not quite as they should be.
PB: Okay, will compile them.
DLM: Yeah, leaving chunks of data and doing refits.
CB: This is where instead of doing chunks of data we can look at the gradients at the optimum with large data, where half the data is pushing this way and other half the other way
DLM: Oh, yeah! TG can your stuff do that?
TG: Yeah, it does. I can give you a gradient per IC, that’s not hard.
CB: How many compounds you have? I think you can make a 2D-interactive plot with the deltas between QM and MM and we can see the chemistries hovering over.
PB: Okay.