Interested in training a potential for an interfacial system with a metal plate and a solution of water and ion on top.

[AbrarFaiyad]

Greetings, everyone. I am currently working on training a potential for Molecular Dynamics (MD) simulations. The system under consideration is expected to resemble the following configuration:

As a beginner in DeepMD, I have attempted to generate multiple datasets using Packmol, followed by running short ab initio MD trajectories to construct the datasets. However, I am encountering challenges with the behavior of hydrogen atoms in water, which appear to exhibit abnormal dynamics.
This is the training script I am using

input.json

Any insights or suggestions on addressing this issue and on how to approach training the potential would be greatly appreciated.

[wanghan-iapcm]

1. You may find some hint for model precision in the link https://docs.deepmodeling.com/projects/deepmd/en/master/troubleshooting/precision.html
2. Usually the short AIMD trajectories are not good training data for deep learning models as they cannot adequately sample application relevant configurations due to limited sampling time and system size. It is recommended to generate training data by active learning or concurrent learning algorithms, as described by https://doi.org/10.1103/PhysRevMaterials.3.023804 and implemented by dpgen and dpgen2.