2022-10-27 Chemical Perception meeting notes

Inferring Torsion Potentials from Scratch

https://docs.google.com/presentation/d/10Vmh95Fyjat1O3H6165sYPVS-gVsVg0MzIH7C2CRchE/edit?usp=sharing

@Tobias Huefner

• Slide: Fitting phase shift using two cosine terms

  • DM - Is this identical to the standard way of fitting the phase shift?

  • TH - Yes, this allows you to fit an arbitrary phase shift but produces less noisy gradients for the optimizer

• Slide - Outlook

  • DM - This is very exciting. I’d be interested to see the results for test molecules you haven’t trained on and for molecules with more diverse chemistry. That generalization is higher priority for me than impropers or 1-4 scaling factors

• CC - for your toy example, your result only includes terms with a phase shift of 0 deg. Were the phase shifts fitted?

  • TH - Yes, the k_{n2} term associated with cos(n x - pi / 2) was fit to be zero, so the phase shift is zero. This is a result of the fit.

• TG - Why does your result not reproduce Parsley exactly?

  • TH - When the hyperparameter lambda is small, I do get the same torsion types as Parsley. The parameter values are different but have the same accuracy as Parsley. I want to test whether these parameters are transferable to the same extent as Parsley.