Page Comparison

Scoping experiments/BespokeFit 2

• Slides:

  View file
  name bespokefit2-update.pdf

• DC – Can’t recall why espaloma linearized harmonic terms

  • YQ – Espaloma rewrote

  • DC – We’re not doing it the same way as espaloma, just fitting force constants equilibrium distances/angles.

  • YQ – Ah, when I did that, I saw numerical issues.

  • DC – Well, we did see numerical issues, so maybe that’s it.

  • YQ – If you write the harmonics as a combination of 2 harmonics and reweight them, then the gradient is easier to handle.

  • DC – And that’s because the numerical values are on very different scales?

  • YQ – Yes, if you take gradient wrt ?, you’ll get a cliffy gradient, but if you do it wrt ? it’s easier.

  • (see recording ~14 mins)

  • YQ – There was a trick that JC came up with where, for negative phases, you can instead make the amplitude negative. There’s a trick to make the gradients smoother.

  • DC – You’re not tuning the gradients, right?

  • YQ – Right, tuning phases generally leads to trouble in optimizations. JC recommends keeping phases and periodicities fixed and just tuning magnitudes.

  • DC – Ok, I think that’s what we do. We wanted to do amino acids so I’ll test on those.

  • YQ – Is this starting with DESRES/SPICE data?

  • DC – Nope, no external datasets, just energies and forces from ML potential for the input molecule. (though the ML potential itself was trained on SPICE)

• JW – How’s the runtime?

  • DC – Unknown - Tom pope is running them so I don’t have a good sense for timing. Right now we’re running MD with a machine learning potential which might take an hourish. But could run with Sage.

• DC – What are thoughts right now on software engineering and future?

  • JW – Can’t commit without lead team feedback. Will depend on how things look. Will bring this up with lead team and relay news back

  • DC – Funding until the end of september, hoping to have some preliminary results before then.

• CC – I like this idea a lot - How do the results compare to bespokefit 1? And to fits using bespokefit-1-style training data?

  • DC – ML potential should be as accurate as fast QM. We’ll need to evaluate how they compare.

• JW – Excited to see how smee performs in this, too. We’d love to know

  • DC – Yes, we’re hoping to evaluate that as well.

  • PB – BW did some tests using smee to fit FFs. How did those go?

  • BW – Using SPICE it took a long time and a lot of memory and didn’t finish. Using Sage training set the final quality wasn’t that great, but I didn’t include torsiondrives initially so I’m trying that now.

  • DC – JH is working on some improvements for memory limitations in smee. Submitting things in batches seems to reduce resource requirements.

  • DC – Interested in understanding why smee struggled with fit quality.

  • PB – When trying smee, I got a lot of force constants being 0.

  • BW – I noticed that too - Not exactly 0 but big decreases in magnitude

  • DC – TP and JH mentioned that as well. Interested to try some of the gradient smoothing stuff from YQ’s work.

  • YQ – Happy to share code, will post paper in channel.

• PB – Will MACE potential for charged species be out soon?

  • DC – MACE doesn’t currently support charged species and has some license limitations. The alternative is AIMNET2 which DOES support charges species and is available for industry use.