Is it possible to have a feedback (reinforcement) loop

[afonari]

The idea is to run MD using model. In the case if uncertainty of the model is high (as evaluated from Gaussian kernel or similar) call backend (using iPI for example) and to perform the actual calculation, and use the results to update the model.

[jameswind]

We are on the way of releasing the code for this paper: "Active learning of uniformly accurate interatomic potentials for materials simulation." *Physical Review Materials* 3.2 (2019): 023804. which will meet your purpose. This would take one or two more months since we are stilling testing it and preparing better documentation. For now, I think you could also realize it in a more manual way: 1) during training you could obtain a few, say 4, different models by changing the seed in the training input file; 2) we have an interface with the LAMMPS code, which prints out the uncertainty, we call model deviation, on the fly. After training, one may print out the model deviation when running MD with LAMMPS. For example, if one has trained 2 models, `graph.00.pb` and `graph.01.pb`, one may use the following script ```bash pair_style deepmd graph.00.pb graph.01.pb 10 model_devi.out pair_coeff ``` Here `graph.00.pb` and `graph.01.pb` are served as a list. Doing so LAMMPS will print the model deviation in terms of energies and forces in the file `model_devi.out`, in every `10` MD steps. 3) definitely, you could do some sampling using i-PI or any other MD engines and test the model deviation when you obtain a bunch of trajectories.

…

On Fri, May 10, 2019 at 2:17 PM Alexandr Fonari ***@***.***> wrote: The idea is to run MD using model. In the case if uncertainty of the model is high (as evaluated from Gaussian kernel or similar) to call backend (using iPI for example) and to perform actual calculation, and use the results to update the model. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <https://github.com/deepmodeling/deepmd-kit/issues/28>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEJ6DC3C3VZDJUH5GVNPR5LPUW34FANCNFSM4HMFHI7Q> .

[afonari]

Thanks a fast reply. I will study the paper.