2025-10-22 Cole Lab Check-In Meeting Notes

CCSD(T) + Basis Sets

DC/ AA

• DC: Alice has been collecting CCSD datasets, thinking of applying this to protein force fields in the future. Thinking about basis sets.

• AA: Previously doing unrestricted CCSD(T). Corrected up to quadruple zeta. Were using energies and forces in PySCF. No one else has ever tried complete basis sets with forces – used correction rather than extreapolation.

• AA: Currently for SPICE, running local coupled clusters. I want to run coupled cluster calculations for SPICE. Because not unrestricted, can go up to triple zeta. Can we do CBS extrapolation for forces? I think no, just correction to quadruple zeta. Double zeta forces terrible, triple better. Any comments?

• AA: Nobody has really rigorously benmarked CBS extrapolation.

• DC: Pavan’s paper helpful.

BespokeFit SMEE Progress

FC

• DC: Is your statement that ESPALOMA CDF has less an RMSE < 0.3 kcal/mol consistent with your assertion that it is worse than Sage 2.2.1 for the torsion example on slide 10?

  • FC: The difference is that on slide 9 we are using YAMMBS which uses OpenMM for minimization, while slide 10 uses geomeTRIC. Also slide 9 is RMSE and slide 10 is RMSD.

• DC: When applying the “early stopping” are the changes compared to Sage?

  • FC: Yes

  • DC: We’ve talked about the regularization in ESPALOMA on impropers and didn’t know why, maybe this is why?

  • FC: Yes it looks like I should look into it more

• DM: I would guess that the early stopping with impropers fixes the “catastrophic failure” resolve to an extent but because in the QM the planarity in the ring changes with the torsion so there won’t be a perfect fit.

  • DC: Am I correct that the improper should be driving the ring to be planar?

  • DM: Yes, Pavan is working on alternative functional forms that could be more dynamic

  • FC: I do think there is better performance that can be squeezed out because of the early stopping improvement

• JC: Is the early stopping from a strict number of epochs or a higher tolerance?

  • FC: I actually manually look at when it starts leveling off and then stops it

• AA: …

• DC: The regularization sounds more automated and higher priority

  • AA: How would you handle that for torsions?

  • CC: You can do …

  • AA: It seems like the issue is with the scans

  • FC: Per mol hyperparameters would be labor intensive but seems necessary because of these larger molecules with catastrophic failures. It seems to be related to how wrong the parameters are.

  • CC: I think if you tune hyper parameters correctly you should be able to find a happy medium.

• JC: You brought up differences between OpenMM and geomeTRIC, I wanted to point out that geomeTRIC is meant to find a QM global optimum so it will tend to just to different conformers so that your population of structures isn’t apples to apples.

  • FC: I do find similar trends between OpenMM and geomeTRIC, but I made my own version of YAMMBS that uses geomeTRIC because OpenMM scans the torsions while freezing all other bonds and angles.

  • DC: Oh, I think I like geomeTRIC better then.