2024-10-28 Science team meeting

Project updates

BW

• Past week

  • Refiltering industry dataset

    • BW: crashed over the weekend, restarted now. It ran out of memory (although error message was that it ran out of walltime)

  • Kicked off new besmarts fit.

    • BW: still running

    • BW: not 100% sure what TG has done, he’s reset values as per his chemper meetings. He’s also reset torsion multiplicities and stuff. Not sure which torsions were affected and in what ways.

    • BW: just training the force field to data now.

  • SERMACS

    • More academic than expected

    • Interesting talk about SAPT – there’s a paper where they used SAPT to generate FF parameters for a whole FF (from Sherrill group)

    • Talked to student from Sherrill group about fragmentation work for SAPT

      • One issue in SAPT is how to fragment – they may often do it manually

    • Met Lori Burns!

• Next week

  • Refiltering industry dataset

  • qcsubmit and bespokefit PRs

    • Fix the issue in bespokefit with generating FB inputs, JW said he would cut release on merge

    • Will fix create_basic_dataset

  • lipidmaps dataset

  • test organometallics dataset

    • ~100 molecule

      • def2 basis sets generally pretty good (preferred ECP), most modern and flexible

        • def2-tzvp, or pd / ppd

        • Recommend diffuse functions for charged systems

      • B3LYP probably fine, but LMI is partial to wb97x-? (d is suggested in strategy document)

      • wb97m-v is what LMI used in previous work. Can’t calculate post-scf properties in psi4, e.g. polarizabilities (doesn’t have analytical gradients)

      • try bp86, wb97x-d

      • Does Psi4 handle ECPs automatically?

        • AMI: yes

Project updates

AMI

• Past week

  • Submitted new NAGL2 timing benchmark

    • 7/10 problematic iodine molecules worked first try, 2/3 worked when AMI changed the guess

  • Got Sage 2.2.1 + S data + TM data fit going

    • Step size was too large in previous fit, so didn’t converge. Tried Goldilocksing steps but either objective function didn’t change too much or fit doesn’t converge

  • Worked on getting useful code onto GitHub/updated on GitHub

  • YAMMBS analysis

    • Found some molecules had large changes in ddE. ~200 mols minimized to slightly different structure, seemed within regular variation. Those had different RMSD and TFD too, corresponding to different geometries.

    • Another group of mols had large ddEs differences, 10-20, where MM structures were basically identical but the ddE changed a lot. This may have been due to the minimum conformer changing

• Next week

  • Benchmarking Sage 2.2.1 + S data + TM data

  • Continue getting code on GitHub

  • Read through VDW fit code/start VDW fit?

  • Maybe resubmit hessian datasets if QCSubmit PR goes through

  • Schedule NAGL2 meeting this week

LW

• Past week:

  • Porting checkmol to SMARTS

  • NAGL release issues and substructure tooling

  • Getting Evaluator working on Kubernetes a bit

  • Agile experiment

• Next week:

  • Work out checkmol licensing issues

  • Working on Kubernetes Evaluator

  • NAGL release

  • Validating Interchange packing and simulation