search

CompPhysVienna / n2p2

n2p2 - A Neural Network Potential Package
https://compphysvienna.github.io/n2p2/
GNU General Public License v3.0
225 stars 83 forks source link

Ethylbenzene potential #154

Open Matheus-Novoa opened 2 years ago

Matheus-Novoa commented 2 years ago

I tested the ethylbenzene potential, present in examples/potentials folder (ethylbenzene_SCAN). I tested with a etylbenzene molecule built in Avogadro, however, i got extrapolation warnings.

I would like to know if, in addition to the molecule itself, the MD simulation settings influence the model performance.

input.nn.txt lammps_input.txt c8h10.txt

Kyvala commented 2 years ago

Hi, n2p2 cares just about the positions of atoms. The only influence of the MD package is how positions are calculated. If you get extrapolation warnings, check first, how much are symmetry function extrapolated. If it is just a small value from MIN/MAX, there is no need for worry. Nevertheless, extrapolations warnings indicate that the n2p2 model was not trained on a sufficient amount of data and you are touching unknown configuration space.