Previously, temperature-dependent properties weren’t being fitted. Now, faster process can fit across multiple temperatures
Result is good compared to TIP3P, but not as good as OPC3
MG: (slide 7/8) You said you thought fluctuations might be due to polarizable water model requiring longer run, but TIP3P also has those fluctuations
WW: Maybe all of them need to be run longer, or maybe it’s not an issue with polarization
MG: (Slide 8) Heat of vaporization--brown curve is TIP3P, we’re very close
MG (slide 8) Looks competitively accurate vs TIP3P, would you agree?
WW: yes, though not as good as OPC3
WW: OPC3 has good dielectric, not sure why
MG: 1/dielectric is what matters
MG: OPC3 has correction to heat of vaporization so that comparison isn’t necessarily comparable
[a bit more discussion around 14 min in the recording]
WW: angle parameter is 111.8 deg, vs expt 104.52. is that a problem?
JW: If you have workflow set up, you could try re-doing the fit but keeping the geometry at the expt
WW: Started with that, but didn’t give good results, high objective function.
WW: Only TIP3P keeps experimental geometry, and it’s the worst model. Historically, usually need to change the geometry
CB: Is experimental geometry for gas phase?
WW: Yes
CB: Do we know it for liquid? OPC3 also has larger angle, maybe the angle is larger in liquid
WW: Will look into it
CB: Did you compare to polarizable water model? I joined late
WW: no, not yet
CB: Wondering if differences from expt could be due to not including induced polarization
WW: iAMOEBA reproduced almost exact T-dependent density. Direct polarizability only, but has multipole expansion
MG: Slope of density vs T graph is same as coeff of thermal expansion being too big. Could consider increasing weight on coeff of thermal expansion at 300K. Not sure why it would be better than what you did, but could be
WW: Yeah, could try that if we want further improvement
WW: If angle is ok, I want to stop with the optimization, I think it’s good enough
JW: When you tried using experimental geometry, optimization objective increased?
WW: Density and dielectric constant were way too high, that’s why the objective was bad, started too high and any change in parameters made it worse
CB: Trying to brainstorm why qualitative difference from expt? Can you compare with iAMOEBA?
WW: doesn’t work with OpenMM, so would be a lot of work
CB: Does iAMOEBA have large angle also? Would be good to identify which differences are due to polarizability
WW: Sharing iAMOEBA direct paper, almost perfectly agrees with expt. AMOEBA mutual is close but not perfect, AMOEBA direct looks similar to my model
MG: Doesn’t look like a direct mapping from polarization type to result, other factors are affecting it
MG: Would be good to know difference between amoeba direct and iamoeba direct
WW: I think iAMOEBA is flexible water
WW: iAMOEBA has smaller water angle, closer to expt
CB: Charge is different also for O
MG: iAMOEBA has no LJ on H, slightly larger O
WW: We don’t have LJ on H, tried but didn’t help
MG: What’s our O sigma?
WW: 3.189 A, smaller than iAMOEBA
WW: iAMOEBA is 14-7, hard to compare
CB: Wonder if you’re in a local minimum in parameter space
WW: Don’t think so, could be another minimum that is less physically meaningful. E.g. could make sigma very small to get better heat of vaporization, but wouldn’t be physically reasonable. But also don’t have any data showing it’s a global minimum, how could I show that?
CB: Start with iAMOEBA parameters, hold frozen except sigma, re-tune sigma (to adjust for functional form diff), then compare to that model. Then, try another optimization with that starting point, to see where it goes
WW: iAMOEBA is trained on intense training set, fits to a bunch of different properties and data, so would have very different objective function
CB: Changing the starting condition (especially to one that might be close to a minimum, given its performance) could just help you locate a second minimum
WW: Given this discussion, do we think the model needs to be tuned more?
MG: A few of the graphs on slide 8 are noisy, need longer runs…when you do FB optimization, is the noisiness reflected as errors in the gradient? Could lead to trouble hitting minimum
WW: Good point, think that’s possible
MG: How many temperatures are you using for training?
WW: 8
MG: Given that thermal expansion, which is noisy, also feeds into density--might make sense to do longer runs for ~3 temperatures to improve
WW: Can try that
CB: Could experiment by training just on quick properties and see if that fixes it