[Chemicalphysicstheory] DL interatomic potential

[Joakim Brorsson]

Hi Elisabeth

This sounds really interesting, unfortunately the article does not seem to be accessible via Chalmers Library.

If anyone manages to get hold of a copy, please share it.

With regards

Joakim
________________________________
From: Chemicalphysicstheory <chemicalphysicstheory-bounces at lists.chalmers.se> on behalf of Elisabeth Dietze <dietze at chalmers.se>
Sent: Thursday, December 1, 2022 2:20:51 PM
To: chemicalphysicstheory at lists.chalmers.se <chemicalphysicstheory at lists.chalmers.se>
Subject: [Chemicalphysicstheory] DL interatomic potential


Dear all,



This week was a new DL interatomic potential published by Chen and Ong: https://www.nature.com/articles/s43588-022-00349-3 and to be honest, they have a very nice github https://github.com/materialsvirtuallab/m3gnet package which I tested out of curiosity and because it works out of the box on my MacBook.



Comparing to some old data of my PhD I concluded: The potential they provide is probably very good for bulk materials but for surfaces and adsorption it isn’t transferable (yet).

For different metals (see attached file), the lattice constant is well reproduced and the total energy reasonable and in the correct order.

Relaxing for each metal the 55,147,309 cuboctahedral nanocluster and taking the slope of the linear fit of E_cluster/size-E_bulk vs. size^(-1/3) gave a completely off result. Thus, not only the difference of the slopes is large, but the order of metals change compared to DFT.

Does one try O adsorption on the (111) facet of the 55 and 147 Au nanoparticle, the O atom relaxes to a position in plane to the metal atoms.



Anyway, I think their project is a really nice example that one can produce out of the box potentials which work on other people laptops within some minutes, and it is very fast. Also doing relaxation and MD on one’s own MacBook is fun. Also it is all open code in Python – so even if they have not documented the output format, one can just check it or address different classes separately.

They have the possibility to retrain their potential, so maybe feeding it some surfaces could help the results.



Best Elisabeth
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.chalmers.se/pipermail/chemicalphysicstheory/attachments/20221201/6cd99500/attachment.html>