The credibility of the model

[lammps-vasp]

What is the upper limit of deepmd kit test error and training error, and how many orders of magnitude can be reached

[DingChangjie]

Hi, by "upper limit" do you mean how precise the DeepMD-kit could be ? Generally speaking, we care more on the error of force than that of energy. Here is a video, showing that the final error of force could reach an order of 10^(-3). However, you should also be careful of an overfitting problem, which appears to gain a small error but weaken the generalization ability remarkably (See this issue)

[lammps-vasp]

Thank you very much

What is the maximum order of magnitude of the accuracy of deepmd kit?

In addition, I'm training with only energy information, no power information.

[DingChangjie]

You're welcome. Actually I'm also a learner in DeepMD, so some of my opinions may not always be correct...

1. Since DeepMD is based on a deep Artificial Neural Network, we quantify the "accuracy" by training error and testing error, which are reported by the program in the "lcurve.out"(column named l2_f_trn, l2_f_test, etc. the meanings are self-explanatory) . So normally the smaller the better, since large errors are skeptical, indicating a poor predicting ability.
   But this really depends on your own system. For my alloy system, the l2_f_test could reach a maximum order of 10^(-3) regardless of overfitting, which is the same as the video said.
2. By "power" do you mean the virial information ? If so, I recommend to try it, at least it really helped me to decrease the training and testing error.