necessity of training to cohesive energies instead of total energies

[Efrem Braun]

Hello,

I was going through the aenet-example-01-TiO2 tutorial, and I was wondering about the necessity of including atomic energies in generate.in. Per http://ann.atomistic.net/documentation/#input-file-example-generate-in-for-tio-sub-2-sub, the actual values for these atomic energies don't seem to be critical (since they can be either isolated atomic energies or average energies), and I was wondering whether I'd be able to just train to total energies instead. I re-ran the tutorial, just changing the two lines in generate.in to set the atomic energies of O and Ti to 0, but when I tried to train a neural net, the RMSE of the training set was 171 meV/atom after 200 epochs, as compared to 7.9 meV/atom when I left the atomic energies as they were.

I wouldn't have thought that this information would be so critical. Identical results could be obtained by adding atomic energies as the bias to the final linear node of the neural net, so I would think aenet would have been able to train the neural net to figure this out on its own.

Could you help clue me in to what I'm missing? Why are inputting atomic energies so critical to the training?

Thank you!

Efrem Braun

[Nong Artrith]

Dear Efrem,

If the reference data set contains structures with different compositions, the total energies vary over a much greater range than the formation energies with respect to some reference state (i.e., the cohesive energy if the reference states are the isolated atoms). Because of these large fluctuations it is much harder to fit a potential to the total energies. However, the relevant information is really contained in the formation energies and the energy difference between different atomic species is mostly constant. In 'generate.in' you can define reasonable atomic energies to be subtracted before the potential fit in order to minimize the fluctuations in the energies.

In principle it should also be possible to fit the total energies directly as the bias (as you point out) can absorb the atomic energy, but training convergence will likely take much longer.

The above also means that the atomic energies are not important in systems containing a single atomic species.

Best,
Nong