Increase energy accuracy

[Achyut-blip]

Hello everyone,

I got the following warning when I was trying to train a ML potential in GPUMD4.0 version

I couldn't increase the accuracy of the energy test/train.

Warning:
There is energy < -100 eV/atom in the data set.
Because we use single precision in NEP training
it means that the reference and calculated energies
might only be accurate up to 1 meV/atom
which can effectively introduce noises.
We suggest you preprocess (using double precision)
your data to make the energies closer to 0.

How should I preprocess to avoid this warning?

Does it mean that my energy data aren't as diverse as required?

No matter how many generation train, my r^2 value gets stuck around 0.94 with a weird test plot as shown in the attached figure below.

Help me solve this warning.

[luowh35]

We suggest you preprocess (using double precision), your data to make the energies closer to 0.

I think the problem may mainly come from the absolute energy reference. It might be helpful to shift all the energy data so that the reference energy is close to zero. This tool may help you achieve that:https://github.com/brucefan1983/GPUMD/tree/master/tools/Analysis_and_Processing/energy-reference-aligner

[luowh35]

Dear @Achyut-blip,
I'm glad to hear that works for you.
Please feel free to ask if you have any further questions.

All the best,
Wenhao :)

[Achyut-blip]

Thank you so much!

I suddenly got another problem.

I has a train.xyz which has configurations from different unit cell and supercell DFT data

I tried to shift it through the code that you suggested but it displayed an error. So I found that it doesn't work if the training data has multiple different sized structures.

Do you think individually shifting it and then combining the shifted train.xyz to make a overall shifted train.xyz to train an interatomic potential for a combined different sized structure would work?

I really appreciate your time and consideration in this regards

[luowh35]

I has a train.xyz which has configurations from different unit cell and supercell DFT data

This will not induce the error, the code is to shift energy per atom for different type of elements.

Do you think individually shifting it and then combining the shifted train.xyz to make a overall shifted train.xyz to train an interatomic potential for a combined different sized structure would work?

I would not recommend shifting it independently and then combining them, unless you make sure exactly the same atomic baselines are used for all subsets. If each file is shifted separately, the fitted baseline energies may be different for different subsets. After combining them, the relative energy reference between different structures may become inconsistent, which can affect the energy fitting.

Feel free to let me know if you got any error.

[luowh35]

And @Achyut-blip, maybe you can also try this code: https://github.com/brucefan1983/GPUMD/blob/master/tools/Analysis_and_Processing/shift_energy_to_zero/shift_energy_to_zero.py

[Achyut-blip]

Thank you so much the last code you suggested worked for me.

I had another problem that I faced after creating a NEP potential.

I am using NEP-cpu interfaced with LAMMPS for calculating other properties of the material that I am studying.

The energy, force test accuracies are excellent with a r^2 values of 0.99 and low RMSE errors in Energy and forces however, when I do the Green-kubo integration using that potential, the HCACF never decays although I run it for 50-200ns NVE production.

I am showing the hcacf and the corresponding thermal conductivity curve,

Please suggest me any remedies that can be applied to solve this issue. Is this the issue with potential or the structure itself?

You have been helpful throughout the issues I faced, I appreciate you for that.

[brucefan1983]

When you have new question, you can also ask here: https://matsci.org/c/gpumd/68

Usually, along with your questions/results, you also need to report:

• GPUMD version you used
• run.in and/or nep.in inputs, and descriptions for your purpose
• The beginning of your model.xyz and descriptions for your simulation model