[Feature Request] Automatic detmermination of the parameter `sel`

[amcadmus]

Summary

The parameter sel is set by the user in the current implementation. For a user who do not have much knowledge on deepmd-kit may not be able to set it correctly. It would be nice to provide the option that the sel can be determined automatically from the training data.

Detailed Description

A possible way of implementation is to go through the training data, and check the maximal number of atoms in the cut-off radius, then automatically set the sel. This optional should be supported by deepmd-kit.

[njzjz]

We may raise RuntimeError if sel is not enough during the training, if users use their settings, since the model will be bad. We may add a --force option if users insist on using the bad sel.

[AnguseZhang]

Question is that do we only need determine sel in training process , rather than MD process?

[jameswind]

We may allow for a slightly larger number than that decided from the training data, but it should not be too large. In MD if the number of neighbors exceeds sel we should at least provide a warning.

…

On Mon, Jun 14, 2021 at 3:26 PM AnguseZhang ***@***.***> wrote: Question is that do we only need determine sel in training process , rather than MD process? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#729 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEJ6DC4LZDX6W5DPAAJM4M3TSWVKZANCNFSM46QLIBLA> .

[denghuilu]

We may allow for a slightly larger number than that decided from the training data, but it should not be too large. In MD if the number of neighbors exceeds sel we should at least provide a warning.
…
On Mon, Jun 14, 2021 at 3:26 PM AnguseZhang @.***> wrote: Question is that do we only need determine sel in training process , rather than MD process? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#729 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEJ6DC4LZDX6W5DPAAJM4M3TSWVKZANCNFSM46QLIBLA .

Do we need to check the sel size at every step? We may need a more efficient way to check the sel size in MD simulation.

[njzjz]

Do we need to check the sel size at every step? We may need a more efficient way to check the sel size in MD simulation.

Is it hard to check during evaluating? Or we can add a new C++ API or LAMMPS fix to check it.