search

isayev / ASE_ANI

ANI-1 neural net potential with python interface (ASE)
MIT License
220 stars 56 forks source link

Training on atomization or absolute energies? #12

Closed proteneer closed 6 years ago

proteneer commented 6 years ago

Another point that wasn't mentioned in the paper is if you guys were training to the total QM energy or the Atomization Energies (since you guys do provide the self-interaction atomic energies). Can you kindly clarify?

Jussmith01 commented 6 years ago

Training is to atomization energies.

proteneer commented 6 years ago

Thanks Justin (for this and the other issue!)

proteneer commented 6 years ago

I need to clarify something else, are you guys normalizing the energies to the unit gaussian? We think Adrian may have mentioned something about this during the talk he gave at our company, but we couldn't find any references to this in the pre-print of the Chemical Science paper.

Jussmith01 commented 6 years ago

Are you you asking: are we normalizing the labels [energies] (after we remove the single atom energies from the total energy)? If so, the answer is no. We actually do no preprocessing on the energies aside from the atomization shift. Once the energies are shifted we train directly to hartrees, which are in the range -4 to 0 Ha (Figure 3C -> https://arxiv.org/pdf/1708.04987v1.pdf) unless something is wrong with the data (or working with larger systems in the training set). Early on in this work I noticed normalizing didn't make a difference so I stopped using it. However, if working with larger molecules in the training set it might be a good idea to start doing it. Does this answer your question?

proteneer commented 6 years ago

Thanks, this is exactly what I was looking for.