Active Learning on Defects

[mstapelberg]

For Primary Damage Simulations we need the ability to say that our MLIPs can model defects in many different configurations.

We have datasets meant for Primary Damage cascades in V and W already computed, including the following structures:

1. Self-Interstitial Atoms
2. Surfaces
3. Liquid Surfaces
4. Di and Tri Vacancies
5. C15, A15, and other non cubic systems, HCP too
6. Diamond structure
7. Short range dimers

Ideally we should include variations of these for compositions that we aim to create primary damage cascade models in.

A workflow I think could be useful is:

1. Use general MLIP (trained with NEB data and our Active Learning) to find compositions with highest PEL heterogeneity
2. Generate additional defect data for those specific compositions, in addition to existing general MLIP data
3. Train Scalable Allegro or MTP on that specific dataset

Work that needs to be done for this:

• [ ]