Skip to content

Numerical instabilities in NativePET #562

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
frostedoyster opened this issue Apr 25, 2025 · 1 comment
Open

Numerical instabilities in NativePET #562

frostedoyster opened this issue Apr 25, 2025 · 1 comment
Labels
PET PET architecture Priority: Medium Important issues to address after high priority.

Comments

@frostedoyster
Copy link
Collaborator

This was observed by @tulga-rdn when using it with long-range effects

@frostedoyster frostedoyster added PET PET architecture Priority: Medium Important issues to address after high priority. labels Apr 25, 2025
@tulga-rdn
Copy link
Collaborator

tulga-rdn commented Apr 29, 2025

It seems that it is due to the metatrain infra (perhaps the composition model), not the specific model. I ran some experiments comparing energies of a molecule with varying degree of perturbation of the atoms. What I observed is:

  1. New metatrain PET/NativePET and nanoPET show the same energy for some rattled molecules. This behaviour disappears when I change the default dtype (base_precision) to float64.
  2. Old metatrain (commit a624a96) nanoPET does not show any "step-wise" energies with both float64 and float32

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
PET PET architecture Priority: Medium Important issues to address after high priority.
Projects
None yet
Development

No branches or pull requests

2 participants