Dataset filtering

• Maat and Jang need to make code to make training data for the next fitting generation

• HJ – Will be based on taking a large dataset of compounds, clustering, and picking the molecule with a lot of coverage

• JM – Open to a lot of different options for how to do this

• JH – Filtering could become a component in the submission workflow

• JM – This dataset would be used for the Sage fit

• JW – Are we approaching a limit on how much data we can feed into ForceBalance?

  • JW + HJ – Let’s assume no upper limit for now

• HJ – We will use Roche set, coverage set, new set

What will dataset look like?

• We have 200 torsion terms in our FF, so we’d want 5 scans for each torsion, so 1000 torsiondrives

QCArchive submission

Timeline

HJ – Running optimization takes ~1 day, sometimes need to run 5 times for data trouble.

JW – Assuming 2 days per torsion, time 1000 torsions.