2021-03-19 QCFractal User Meeting notes

QCArchive status

Ben

• Update failed on production, public instance to latest QCFractal server release

  • there are database inconsistencies that blocked this

    • duplicate molecules; molecules are uniquely identified by a hash, but we actually have duplicates (same hash, same everything)

    • some knock-on effects: likely duplicate optimizations, duplicate torsiondrives

      • id of the molecule is built into the hash for the procedure

    • second inconsistency: duplicate SingleResults (energies, gradients); null columns in MM cases cause deduplication components to not work reliably on these

• Not so easy to fix, because duplicate cases often have foreign keys in JSON blobs

• What are the possible solutions?

  • Migration of old database data via server into a new server instance

    • Molecules, collections, etc.

    • this approach could be designed to preserve internal IDs where possible

      • SB: I use IDs to refer to data, since that’s the only reliable way to reference exactly the same data over time

• SB: one thing that’s missing is a data retention policy that demarcates expectations around data preservation, even through a migration

  • BP: this is something we’re currently thinking on

• SB: is it possible to specify that I want to create a new dataset with specific record IDs instead of relying on deduplication of resubmission of same data.

  • BP: not really possible at the moment; only possible to specify existing molecules and specs that go into a collection, not existing results

• HP: how are molecules saved in the database? Is it just XYZ coords, or other information?

  • BP: The hash function used to determine identity includes several fields in the QCElemental Molecule: https://qcelemental.readthedocs.io/en/latest/model_molecule.html#molecular-hash

Dataset status

David

• SB: on WBO TorsionDrive Dataset

  • don’t see all of the original molecules that I expected to see from the source set

  • at the very least the CMILES aren’t present as expected?

  • only tautomers and protomers are there, not the base forms

  • JH: QCSubmit is dropping the input for the enumeration stages

    • the toolkit won’t return the input molecule, and QCSubmit only taking what comes back from the toolkit

• DD: public datasets prepared, putting finishing touches on READMEs, adding dataset PDFs

  • TG: perhaps consider waiting for new basis set if it’s coming soon; there are iodines in some of the datasets

  • SB: autoaux is the one we’re likely to choose; 15% performance decrease, but should give the same results for non-iodine

  • TG: there may be some new basis set names

  • SB: would like to get something fairly soon, so perhaps submit at least one of these upfront?

    • DD: can submit the Swope set with existing basis set

  • [decision] submit Swope set first, will give us starting data, help identify issues in the process

User questions

• HP: Seeing SCF failure in my own datasets, planning to increase scf_maxiter, changing the basis set. Are there other changes I can try?

  • PB: soscf in psi4, turning this on will enable second order consideration during SCF convergence