Predicting forces

[Hasan Babaei]

Hello,

I was wondering if aenet trains also to forces.
I have generated an NN potential using aenet which predicts energy of a system perfectly (within 0,4 meV) but the forces are too far from what I get using DFT. During training, RMSE and MAE converges very well. Thanks in advance for you help.

Best,
Hasan

[Nong Artrith]

Dear Hasan,

Unfortunately, the current public version of aenet does not yet support force training.

Best,
Nong

[Hasan Babaei]

Thanks for the quick response. I am wondering how it predicts forces though?
Hasan

[Hasan Babaei]

And it also reads forces when training.

On Monday, July 23, 2018 at 4:33:17 PM UTC-4, Nong Artrith wrote:

[Nongnuch Artrith]

The atomic forces are computed from the analytical derivates of the ANN potential.  Please take a look at the aenet source code, and the expression is also given in equation (11) in Artrith and Behler, Phys. Rev. B 85, 045439, 2012 (https://doi.org/10.1103/PhysRevB.85.045439).

Note that the quality of the forces generally gets better the more reference data is used for the potential fit.

Nong

[jia...@vt.edu]

Dear Nong,

I have one follow up questions. If training the ANN potential only by the energy, can this ANN potential be used as molecular dynamics calculator? Since the atomic force is just derivation from the atomic energy, so is it reliable to run the molecular dynamics? 

Best,

Jiamin

[Nongnuch Artrith]

Dear Jiamin,

Yes, you can use aenet's ANN potentials in your MD simulations.

Please confirm the accuracy of the atomic forces by comparing the predicted forces with your reference data.  If needed, the forces can be improved by including additional reference structures in the training set.

All the best,
Nong

[Jiamin Wang]

Dear Nong,

Thanks a lot for responding. I will try it.

Best wishes,

Jiamin

--

Jiamin Wang

Ph.D candidate

Department of Chemical Engineering

245 Goodwin Hall, Virginia Tech