2021-09-24 QC meeting notes

Updates from MolSSI

Ben

• PR with prod server fixes is up, not yet merged

• want to pin QCEngine, QCElemental for now on both current generation of QCFractal and next branch

  • new QCEngine, QCElemental should be backward compatible, but want to move cautiously for a while and advance next

• putting together job posting for new postdoc, targeting AI/ML applications with QCFractal

New submissions

• CC: was able to solve issue with proline last week, had to handle it with some nasty geometry work, but should be good enough to use

  • even in forcebalance paper with scans for proline, not clear they actually did any

  • only other hurdle: need to ask for a constraint on the dihedrals that are not being scanned in QCSubmit

    • want to constrain sidechain dihedrals during torsiondrive

    • JH: might need to modify QCSubmit a bit; we haven’t had to do this before

      • CC: can put together an example as an issue on openff-qcsubmit

• PB: Single Points #227; any update on errors/progress?

  • DD: error cycling consistently chokes on it; getting persistent ConnectionRefusedErrors due to query timeouts on server

    • tried reducing batch size of individual queries to alleviate; added retry logic; still no dependable success

  • PB: for this dataset, do we know why the submit response is 4496?

    • JH: for the dispersion correction, the server does add additional calculations; may be coming from there

• DD: JH, we’ve put your new submission in and it should be going through error cycling. I’ll be spinning up XTB and ANI workers.

  • JH: Sounds good. I have access to Newcastle compute cluster now so I can spin some up too. I could use any docs about getting this running.

  • DD: Can do, PB and I just worked on this the other day. We used the benchmarking instructions to get the compute spun up.

    • (JH and DD will work on this after the call)

User questions

• WW – I’m using psi4, doing 7 calculations each run, 6 with imposed electric field in cartesian directions, and one with no electric field. So I’m looking for openFF datasets with ESP calculated, and then I’ll expand those with additional external E-field calcs.

  • I’ll use Hyesu’s RESPYTE to process the output.

• LW – Does RESPYTE use MSK shells? I thought it used different radii

  • WW – I do, but Hyesu’s package specified different radii

• DD – So you can get wavefunction from completed jobs, then use RESPYTE to process it, then use Simon’s tools from the slack thread to get the ESP.

• DD – Resources for making QC dataset submissions:

  • See STANDARDS on qca-dataset-submission

  • Example for dataset generation:

• WW – I already have molecules ready, is input SDF?

  • DD – The input can be a variety of formats. SDF is one option.

  • WW – I’d like to use existing molecules/MP2 results, where the conformations are meaningful, so I’ll use SDF as the starting point.

• [to do] write up README quickstart for new users

• (DD and WW will schedule a working session to start submission)

Science support

• DD – We’ve been reviewing/gathering upcoming plans for next year: Quantum Chemistry Infrastructure / QC*

• JH: an officially-supported way to build the ESP from the wavefunction in psi4, that can then be used for e.g. optimizations as the initial guess

  • I think psi4 can already do this with its own format; but support is needed for passing in and out of it

• CC: another area of interest: being able to create a single submission that chains two methods, e.g. one cheap and one expensive, for an optimization

  • BP – Agree, this could be a service that runs on QCFractal. I’ve thought about this on-and-off. For quantum chemists its common to want to do an optimization and then a property calculation.

  • JH – This would be immediately applicable to Hessian calcs.

• On Rosemary:

  • WW: CB says polarizability support important, and I would want to use the same dataset with polarizability fit calcs.

  • JW: at the moment, we don’t really know what the Rosemary datasets look like; Chapin is investigating this; electric field calculations likely involved, like those you are doing

    • we do want to start experimenting before we have datasets intended for Rosemary training

• CC: some way to account for solvent probably key

  • don’t think there are infrastructure needs

  • can do what Mike Schauperl did with a gas phase calculation, solvent

• JW: from what I understand, PCM is single-core, and really slow

  • perhaps improving this would be high-value for us in Rosemary

  • BP: a lot of solvent models come in through an external library, so not clear what’s required to improve performance

  • CC: have some ideas beyond the scope of this meeting

  • JW: we do have a conversation with Lori Burns on psi4 for conda-forge; might be something we can add to the agenda for that if we anticipate it being very key

• JW: think a reactive stance is best we can do at this time; as Chapin encounters issues, we can address

  • CC: may also hash some of these out in the roadmap update