2020-12-04 Meeting notes

30 min

Trevor Gokey

Parameter splitting progress and follow up from last meeting

• Recap of bit representation strategy for representing force field parameters

• slide 1: parameter splitting on real FF example

  • start with simple FF (1 parameter?)

• slide 2: first example, turning off the hydrogen bit

  • started with the one that had biggest contribution to gradient

  • high regularization penalty

  • doubts that it is realistic that CC and CO bonds are identical (colored orange)

  • further splits split off different chemistry

  • suggestion to parallelize a bit to speed up process

  • general process is to run FB, split and run FB again

• slide 3:

  • unclear how gradient magnitude is affected by splitting

  • regularization: L2 reg from FB

  • L2 reg may not be in the obj and its grad, so that might explain strange behaviour of gradient along splitting