2024-09-09 Science team meeting

Project updates

LW

• Genentech

  • Project proposal

• Shirts

• Polymer simulation [?]

• 2.2.1 full release

• Virtual sites

  • fitting vdw terms with smee and experimenting with scale factors

Project updates

BW

• Torsion fitting / restrained minimization

  • Background:

    • https://openforcefieldgroup.slack.com/archives/C8P54J2JY/p1723762239389589

    • https://openforcefieldgroup.slack.com/archives/C8P54J2JY/p1723762509503209

  • BW: looked at restraining the minimization with a torsion force

    • LW: has the contribution from the angle force reduced now that angles are allowed to relax?

    • BW:

      • port the code changes back and compute the force contributions

      • figure which scale factors for the torsion restraint give the best balance between minimal overall RMSD changes from the minimized QM, vs allowing really bad forces to relax a bit, also vs keeping torsion within 3 or 5 degrees of the required grid point

      • Use a dataset of 10

• smee

  • benchmarked at 350 iterations

  • BW: want to look into batching with the training

  • LW: maybe try a SPICE run with scale factors set to produce force constants and angles around 1-ish

    • bonds ~500, angles ks ~100? for 2 weeks

    • smee-opt-td if that’s easiest

  • LW: did the gradient sign affect your work?

    • BW: had another sign flip that cancelled out

    •  

• datasets

  • lipids dataset is with Shirts group

  • sulfurs dataset is ready, will send PDF to LW

    • AMI: did you grab sulfur with 4 single bonds? Probably best as a separate dataset. Expected to be a seesaw shape with lone pairs

• chemical perception work

  • BW benchmarked

    • besmarts FF neither converged, nor ran out of steps, nor crashed

    • Using 2.1 priors

•  

Project updates

AMI

• NAGL2

  • AMI: looked into diffuse functions with PCM. Was able to get DDX to work if vdW scaled up. Typically Bondi scaled up by 1.2x, typically UFF by 1.1x, but also is pretty interchangeable.

  • Psi4 defines area of integration sphere in A, but PCMSolver defines it in bohr. Could be an issue.

  • LW: could you please mention that on GitHub in case that is the problem?

• Dataset test set

• Next week:

  • Getting NAGL up and running

  • Parameter splits