Shift in energy test set and pressure

[cesaremalosso]

Summary

Shift in energy test set and pressure when I run it on a dataset with different number of atoms with respect to the training one.

DeePMD-kit Version

v2.1.1.dev80+g8c58455

TensorFlow Version

2.3.0

Python Version, CUDA Version, GCC Version, LAMMPS Version, etc

No response

Details

Hi, I trained a NN over a data set made of snapshots a 432 atoms system. When I test this NN over a different dataset made of 576 atoms ( of course not included in the train one) I find good predictions for forces, so the dynamics of the system should be fine, but I find a rigid shift upon the total energy and also on the diagonal elements of the viral.

energy

yy diagonal element of the virial

On the training set these shifts are not present. I thought that specifying the atomic energies of each species would have solve the problem, but It did not. Is there any kind of solution to this problem?

Thanks

[njzjz]

Do your two datasets have the same chemical composition?

[cesaremalosso]

Yeah, it is NH3

[wanghan-iapcm]

The shift in virial is not expected. Could you please check, by your labeling method, if the virial is equal to the derivative of the energy?

[cesaremalosso]

Do you mean that you want me to check if the first-principles calculations (made with quantum espresso) are correct? I've already checked the convergence of the viral tensor and it's converged with respect to the plane waves cut off that I used. I've also noticed that training on the virial seems to reduce this shift but I was wondering why is it a rigid shift and not just a random error of the NN. It seems like there is some kind of bias somewhere.

[wanghan-iapcm]

You may check if the following equation holds

$$P_{\alpha\beta} = - \frac{1}{\det(h)} \sum_{\gamma=1}^3 h_{\beta\gamma} \frac{\partial E}{\partial h_{\alpha\gamma}}$$

where $h_{\alpha\gamma}$ is the $\gamma$-th component of the $\alpha$-th cell vector.

[njzjz]

Can you provide an example for reproducing?

[cesaremalosso]

Here's the link
In the link you find a directory that contains three different directories:

• model_training: contains the model and the training dataset (generated with DPGEN) and a directory called test_dataset that contains data taken from an ab-initio MD simulation (runned with quantumESPRESSO);
• scf_qe: contains an example of a first-principle calculation (quantumESPRESSO);
• dp_test contains the test over the test_dataset;

[njzjz]

I plot dp_test/test.e.out and obtain a different figure:

In this figure, the bias is about 5 eV for 576 atoms instead of 15 eV in your figure.

Note that your training data has a similar bias.

[cesaremalosso]

Yes, you are right, the model that I linked you is trained over 2 million steps while the one of the plots over 1 million steps. I LINK you the correct model.
aGAIN:

• model_training: contains the model and the training dataset (generated with DPGEN) and a directory called test_dataset that contains data taken from an ab-initio MD simulation (runned with quantumESPRESSO);
• scf_qe: contains an example of a first-principle calculation (quantumESPRESSO);
• dp_test contains the test over the test_dataset and over some of the training data .

You can see that theres is a shift both in energy and virial (diagonal) in the test_data while there's not in the train_data. Is this only a matter of the quality of the training, in the sense that increasing the number of training steps would solve the problem? Why is the error such a rigid shift in the diagonal terms of the virial? Why are the off-diagonal element of the virial instead correct?