2020-08-05 System Object Meeting notes

Owned by Matt Thompson
Aug 05, 2020
1 min read

Date

Aug 5, 2020

Participants

  • @Matt Thompson

  • @Simon Boothroyd

  • @Jeffrey Wagner

Goals

  •  

Discussion topics

Pick one!

Item

Notes

Item

Notes

Interoperability

  • Arguably alone (i.e. without enabling improved optimization or other features) doesn’t really justify the work of the System object

Friction with other infrastructure

Relationships with Toolkit Classes [WIP]

Optimization (force field fitting)

  • JAX example updates

  • Scoping out how optimization will work long-term with larger datasets

Data management (reproducibility/provenance)



Jeff drew up matrix diagrams (P, Q, M, L), took notes here What we made today: https://docs.google.com/presentation/d/13r7MLJ6tr2kLoeCBGKP7U7lsa2zgJDLt3QFpQDuifzQ/edit#slide=id.g8e4974814e_0_113

Action items

  • What assumptions have to be true for VJP to be pre-computable? 

  • What features of angles make it satisfy this requirement? 

  • What other functional forms for angles or other things WOULDN’T work?

  • Will VirtualSites (where geometry is also dependent on FF params) cause huge problems?

    • Lead to a need for inner and outer loops where different sorts of parameters update in each?

  • Do we expect to do “convex optimization tricks”, where we arrive at the final answer using just one big matrix inversion? Did Jeff just dream this up altogether? He thinks something like this was used in lsfitpar.

Decisions