2021-05-19 MT/JW Jax Handlers check in

Participants

  • @Matt Thompson

  • @Jeffrey Wagner

Goals

  • Decide a course of action + timeline for interfacing OpenFF parameterization with timemachine

Discussion topics

Item

Notes

Item

Notes

Options for timemachine integration

  • MT – Three major routes:

    • Something one-off with the toolkit/system (system.handler.to_ff_params and to_system_params)

    • Something like smirnoffee

    • Something like what TG is doing, with parameter vectorization

  • JW – Maybe the best thing we can offer is a test suite for “here are some reference inputs and outputs to ensure that you’re assigning parameters correctly”

    • Maybe also a FF spec downgrader? So that, if we start changing the spec version, it can be exported to a timemachine-friendly format

  • What about integrating with TimeMachine via replacement of handler.parameterize?

    • Con: Custom implementation, unknown responsibility for maintenance

    • Con: There’s already an implementation to handle all current terms in timeachine

    • Pro: Would cover eg. WBO torsions the day that it gets in a main-line FF

  • When we start doing protein stuff, JW could add the necessary performance improvements (like caching here)

  • Can audit parts of timemachine code where differences may be emerging – Eg AM1 restraints etc.



Decision: We won’t try to force new things into TimeMachine code in the next 3 months, but we will audit it when we find bugs in the toolkit and actively propagate resolutions into their code where possible. We will also consider creating a “reference energy” dataset which can be widely distributed for others to validate their implementations.

Action items

Decisions