2021-06-09 Bespoke workflow meeting notes

Alternative functional form plans

• DC – Shows slides

  • Link slides here

• DC – (Testing: Water Model slide) Not sure how to correct for constrained h bonds in heat capacity. Could use LPW advice

  • LPW – When classical models are used to calculate heat capacity, the results are always way off from experiment, especially with high frequency degrees of freedom. This is because of quantization effects. But whenwe use a rigid water model, a tiny amount is still thermally accessible. So the correction that we apply is that we take 7 experimental IR peaks of liquid water, and then we subtract the contribution from the FF, then add back the contribution from a quantum harmonic oscillator. In practice, this means adding a constant that depends on temperature. But those numbers are all stored in the text tables in forcebalance. So if you’re looking a tthe fluctuation in the instantanous enthalpy, you’ll need to add those numbers manually.

  • DC – In the middle column, I’m using forcebalance to do a one-off calculation

  • LPW – Oh, then that’s worth looking further into.

  • DC – I’ll send the code

• DC – Could we use openff-evaluator to get a load of hydrocarbon liquid properties? I suspect the answer is yes.

  • LPW – One thing I found when I was working on Parsley is that the ThermoML database does not have a lot of older physical property measurements with simple molecules (like ethanol). So there may be some need to do manual inputting of literature data.

  • SB – Agree, thermoml lacks simple molecules. Though it does have refrigerants and long-chain alkanes. So it’s worth looking there to see whether it has suitable data for your use cases.

  • LPW – Re hydrocarbons: I haven’t tried to build a hydrocarbon FF before but what I’ve heard is that linear dependencies in vdW parameters can be fairly troublesome (C and H radii can move opposite to each other and give very similar results). So this may raise a question about temperature ranges and specific data. It may also help to have halogens on there as well to hold things stable.

  • DC – Also, maybe pinning to some QM-derived values would also help avoid this linear dependence.

• On “Parameters from QM” slide

  • LPW – Re potential form – Are there some concerns that combining exponential decay with this form would lead to some undesired bumpiness in the energy landscape

  • DC – I’ll have to look at this.

  • DM – Chodera would also ask that we ensure that there are no singularities in this functional form, because otherwise it won’t be suitable for some free energy calcs

  • DC – This form should smoothly go to 0, though I could see this having trouble for other reasons in free energy calcs. I’ll see what this looks like with a bunch of lambdas

• (Post slides)

  • LPW – On the last slide, my thought from the outside is that, if we want to think about bespoke vs. transferrable… transferability is often considered a “holy grail”, but it’s very hard to define what it is. If this was able to demonstrate transferability that would be great, but having said that, I don’t know if it’s possible to do that in 3 months. Though a limited example could be good.

  • DC – I don’t think we should feel constrained by “3 months”. CR has a lot of experience as a developer of qubekit, and JH will be around for a long time as well.

• DM – Chris Rowley is fairly accessible, and is interested in OpenFF. The limiting factor with getting him involved was support for alternative functional forms. So if this is a good proof of concept, he’ll be interested in this as well.

• JH – For C6 and C8 coefficients, Rowley has derived guesses for initial values for those from GAFF. So we might consider doing the same.

  • DC – Agree. That could be a good first approach, while we work on more detailed approaches.

• DC – Is everything in place from the SMIRNOFF plugins side? SB and JH were seeing some potential issues with vsites and forcebalance

• LPW – Do you think B68 keeps values tied more closely to QM?

  • DC – I think the paper was more of a proof-of-principle. They did radial distribution functions, but it doesn’t seem like they went far past what LPW did before

  • LPW-- Is there some flexibility in the model that wasn’t exercised in the water study that might be needed in solution properties. Like if you compare B68 to TIP4P, B68 has this extra parameter. If it then does equally as well in fitting experimental properties of water, then the extra parameter may be useful to fit something else in some systems. Do we have any idea what this parameter looks like or where it would start making a difference? Maybe with ion interactions or other solution properties? This could be an interesting avenue of research.

  • DC – That is interesting. Are there other solution properties that don’t require free energies?

  • SB – Mixture enthalpies are available. These are available in the Sage training set.

  • DM – We’re enthusiastic about mixture data, since it doesn’t require gas-condensed phase transfer, so we don’t have to worry about solvation effects in charges/interactions.

• DC – So, for next steps, it’s fairly clear what we need to do

  • Getting a liquid property database for hydrocarbons

  • Getting QM data (initial guesses?) from literature and figuring out a general way to derive them.

  • Fixing heat capacity issue

• DM – May be good to look at Phil Huenenberger’s work, on chlorohydrocarbons.

  • SB – I beleive that all of PH’s data in in ThermoML