Skip to end of metadata
Go to start of metadata

You are viewing an old version of this page. View the current version.

Compare with Current View Page History

Version 1 Current »

Date

Participants

Discussion topics

Notes

  • JH – Worked mostly on QCSubmit. Working on speed and compression with TG.

  • JH – Still having convergence problems with ANI. Getting data together to submit to Adrian Roitberg

  • JH – Using XTB at the moment. Has high element/chemistry coverage and speed. Running on JACS torsion fragments. Considering using this as a supported option for bespoke workflow, so currently benchmarking time/performance.

  • JH – Thinking about benchmarking for bespoke workflow. Will generate lots of conformers at high temperature. Then will rank energies before and after fitting torsions.

  • JH – Another option for testing workflow would be to benchmark against existing JACS torsiondrives. However this comparison won’t include ANI.

  • JW – What benchmarking strategies are available? QM minimum comparison, high-temp conformer comparison, phys props for small molecules, host-guest systems.

    • DM – host molecules torsions have been identified as a problem. So refitting might improve performance.

      • Hosts are “acd” and “bcd” –

    • JH – Could also test against freesolv

      • DM – I don’t think torsions would make a big difference on these.

  • Benchmarking infrastructure

    • (General) We probably can NOT use QCA/QCEngine as the backend for the industry benchmarking, since it won’t allow input of custom OFFXMLs.

  • Pre-release benchmarking

    • JW – If this can be run in less than ~an hour, we can automate it using GitHub actions.

    • DM – It’d be good for this to be available for manual runs, so that individuals iterating on new FFs can quickly evaluate their changes.

  • JW – JH, please (re)send the code where you deduplicate redundant SMARTS

  • PB – To confirm, new parameter testing is currently done by manually re-running benchmarking.

    • DM – Correct. For released FFs, we can use QCArchive/QCEngine. For unreleased FFs, we use benchmarkFF.

    • (General) Added some thoughts to upcoming FF release meeting agenda item 2020-10-15 Meeting notes

  • JA – Are we thinking of having a fixed test set?

    • DM – It will probably be fixed for set periods of time, but may be improved/updated roughly annually so that we can ensure it’s meaningful and not overtrain

    • JA – Can we do cross-validation?

    • JW – Probably not, current fits take about 2 weeks, so CV folds would make FF fits take several months.


Action items

  •  

Decisions

  • No labels