Item | Notes |
|---|
Josh’s last day at newcastle tomorrow | DC – JH is having his last day at newcastle today. He’ll be moving on to OpenFE JW – Still up to own bespokefit, or should we discuss that with OpenFE? JH – I’m happy to keep that. DC – Who should own this calendar meeting when JH changes positions? I’ll hold on to it until things change. |
eyeballing charges coming out of new charge model | https://docs.google.com/presentation/d/1EM_bWLj8n9dSNJxVVLor7tLs-uIGNgA2jGezUNBeIsA/edit#slide=id.p DC – Pyridine nitrogen is strikingly different between AM1BCC and all QM. JR – Might be that the nitrogen needs some description of lone pair. Is there an experimentally measured dipole? DC – There probably is, we should look into that. CA – JH, as you mentioned, much of our training set was druglike fragments which are larger than this. JR – Right, but QM should match experiment as a first check.
JW – I see in the QM there are some different charges assigned to graph-symmetric atoms. Worth mentioning that for OpenFF charge models we’d want graph-symmetric charge assignment DC – We’ve discussed with LM and LW about whether we could train a LJ model on these charges. Don’t have to decide now but we’re excited about this direction. Also I’ve heard that halfway between solvent and vacuum charges is ideal. And we might want to include that scaling parameter in an FB fit. BS – Water dipoles seem a little off from what I’m used to. DC – Yeah, we might treat water specially JR – But we want it to be semi-polarized in this case, since it’s going into the FF BS – Would water charges go directly into FF, or would they be averaged with gas-phase? DC – JR – … (see recording ~20 mins) DC – Our best model for water comes out around 2.5 debye, which is a bit low in water. This is a known issue because of missing hbonding. This is why ipolq method would be a step up. JR – I looked into MBIS methods and talked to authors. They said it was a better model of electrostatic energy closer in. But you have to acknowledge that you doin’t get penetraiton energy or first order exchange. But it’s pretty good overall, and it converges at lower order than things like DMA. DC – Compute-time-permitting, I’d be in favor of throwing those into FF training and seeing what comes out.
|
update on smee-valence accuracy
| https://docs.google.com/presentation/d/1MXauooJ-MZjrVzjfL7XEs45vLEC7if7CoUdwsRgsf0E/edit#slide=id.p DC – Last time we left off, we said that training valence parameters to SPICE using smee on a GPU was possible, but we hadn’t benchmarked. JR – The “matched minima” metric? JH – Energy difference must be within 1 kcal/mol, and must be within 20deg of QM minimum JR – is 1 kcal/mol too small? JH – I’m only looking at minima that are within 4 kcal/mol of the global minimum
JR – Do these energy differences between minima correspond to the same values in QM? Rephrased: You determine minima on the QM surface and there are energy differences between minima. You get the same energy difference between minima on classical surface? It’s imporant to get that energy difference right. DC – QM minima are reoptimized with MM and then energies are compared. JR – So this is more about relative energies of minima than the barriers between them JH – Yeah, we could make this more complex and add that, but for now it’s just minima.
DC – Could be worth considering what results would be required for smee to replace FB in OpenFF.
|
seeing as xtb-python is no longer being maintained, is there another fast QM method+package that we should consider for BespokeFit’s “fast” backend? | JH – AIMNET2 is great. That’s what we’re currently using for PL benchmarks. Very fast and high convergence rate. It’s in QCEngine, there’s the pyaimnet2 package on C-F. Licenseing should be MIT. JR – Do we know that it covers all functional groups? JH – That’s kinda a different issue to XTB. XTB works everywhere but isn’t super accurate. AIMNET2 doesn’t work everywhere but is accurate when it does work. JR – When you do a PL binding energy the ligand energy matters a lot. DM – I think we’re OK with incomplete coverage, we can kick people back to full QM.
DC – https://tblite.readthedocs.io/en/latest/users/xtb.html
|
adding torsion benchmark to yammbs? | |