Nep cutoff

[tomasfbouvier]

Hi,

In my nep.txt file I observe that next to the cutoffs there is a couple of number e.g

cutoff 5 5 210 210

I believe that this 210 is the maximum number of neighbours an atom can have. Is this number gonna determine any parameter during the fitting or is it just informative?

[brucefan1983]

The data in this line are

cutoff r_c^R  r_c^A  MN^R  MN^A

where

• r_c^R is the cutoff for the radial descriptor components
• r_c^A is the cutoff for the angular descriptor components
• MN^R is the maximum number of neighbor atoms within r_c^R according to the training dataset
• MN^A is the maximum number of neighbor atoms within r_c^A according to the training dataset

The values for MN^R and MN^A should be consistent with the screen output when you ran nep to train the model.

Based on your data, it seems you have very dense structures (perhaps with high compresive pressure) in your training dataset such that the maximum number of neighbors is as high as 210 even though the cutoff is only 5 angstrom.

The values for MN^R and MN^A do not determine any parameters during the training process but are recorded in nep.txt to facilitate reasonable memory allocation. The code will set up a neighbor list by increasing these values by a margin of 25%. If the MD run collapses, it usually indicates that the system has blown up.

If you ran with a single GPU, there should be an output file neighbor.out containing information on the neighbor counts.

[tomasfbouvier]

Understood. Thank you very much. I am only training a nep model, not running any simulations with gpumd