2021-03-12 QCA Submission Meeting notes

Iodine

Trevor

• TG: current thinking: no iodine

  • basis set is fundamentally invalid for Rb and higher

  • Lee-Ping thought we could get away with using ECP

  • but something is missing, and gradients for ECP must be calculated, which is numerically expensive (iterative solver)

  • we’re currently stuck where we began;

  • Susi Lehtola in psi4 community has code that can automatically generate the fitting basis for dzvp for all electron up to iodine

    • but needs testing cycles by Lee-Ping before we can adopt

    • Lee-Ping is engaging directly with Susi, Lori on psi4 slack, working on a path forward

• SB: is current plan for iodine to switch to the higher basis set being developed by psi4; using autoaux approach?

  • also, iodine not high priority at the moment, so okay if it’s not included in our fitting

• DD: there is iodine in our FF params yes?

  • TG: there is one iodine-specific parameter

• DD: is Lee-Ping’s goal to arrive at a new basis set then?

  • TG: I think that’s the impression; this has to end up in mainline psi4, basis-set-exchange

• BP: one thing I’ve wanted is to be able to use basis-set-exchange inside QCEngine

  • say e.g. bse:auto

  • qcschema can specify basis set you’re using, coefficients per element, per atom

• JH: with a normal psi4 run, can we get this with just a keyword?

• BP: Yes, but thinking if we lose something with that path

• TG: Kinda breaks the model?

• BP: Your choice of DF basis theoretically doesn’t change the result…well, guess it is kinda like your model

• JH: the keywords are part of the hash, yes?

  • BP: yes, but fyi we’re moving away from keyword hashing

• BP: have a PR that might speed up calculation submission (I hope)

  • added an index to column in the database that needed to be searched for every single compute looking for duplicate

  • should now be fast due to hash index in DB

  • DD: should be able to submit #160 finally, even without multiprocessing/multithreading client submission

Test QCA server

Ben

• test server is almost up; will notify when it’s in place

• also new release incoming for QCFractal
  

Genentech set 3

Josh

• Touching base with Pavan next week; fragmenter issues

• SB: might be easier to do just a complete rewrite of fragmenter, from what it looks like

  • JH: in talks with Danny Cole for benchmarking fragmentation methods; this could fit the bill for determining if a rewrite diverges in results from the original

  • Also, have a big dataset (benchmark ligands) with 480 fragments that could be a test case; basically, re-create it and observe differences

• SB: how much does tautomer enumeration factor into fragmentation

  • JH: before qcsubmit used to do much itself, used fragmenter for tautomer enumeration, but now that’s done outside of fragmenter

    • fragmenter use limited to fragmentation itself, capping

• DD: will touch base with Jeff on setting up a sprint on fragmenter rewrite; needs to be soon (next few weeks)

Other upcoming datasets

Simon

• will have priority datasets coming through from Hyesu for next FF fit

• JH: not sure we’ve ever done a Hessian dataset with QCSubmit

  • may want to do a test of this earlier with a smaller dataset

• SB: is there still a plan for a Hessian dataset? Where we run an optimization, then a single point calculation at the end to get the Hessian? May require a service on the server like torsiondrives

• JH: may be able to handle this externally to QCFractal if there isn’t any movement on it in the project itself

  • could make this a GHA that operates on a labeled optimization dataset, submits a corresponding hessian dataset that uses the final mols as the objects of the point calculations. Runs repeatedly to get additions to the dataset as optimizations complete in the source dataset.

• SB: MolSSI hasn’t raised any concerns about wavefunction storage, have they?

  • TG: think it’s okay for single point calculations; we’re also only storing orbitals and eigenvalues, so lightest-weight possible

• DD: could turn around GHA fairly fast; need a test set to work on

  • SB: Let’s get one from Hyesu; she can probably turn one around quickly that would also be useful for her beyond just testing the approach