MTP Transferability to Varying Stoichiometries

[evansjc]

Hello,

I’m interested in studying how variations in stoichiometry affect the electronic properties of amorphous materials using MTP (Moment Tensor Potential) models. Specifically, I’d like to understand how transferable an MTP is when trained only on one stoichiometry.

If an MTP is trained on both crystalline and amorphous configurations of a metal oxide with a fixed stoichiometry (e.g., matching a known crystalline phase), how well can it handle variations in stoichiometry? For example, in a ternary metal oxide (with three metal species), would such a model remain reliable if we change the relative ratios of the metal atoms? If so, to what extent can these ratios be modified before the MTP becomes unreliable?

Similarly, in a binary metal oxide trained only on a fixed metal-to-oxygen ratio, is it reasonable to expect accurate results when varying the metal-to-oxygen ratio outside the training set?

Any insights or best practices for ensuring transferability in this context would be appreciated.

Thank you!

[Anders Blom]

I think it can handle this relatively well, since the stoichiometry itself is not a design parameter for the MTP. What matters, generally speaking, is that the training data for the original MTP contains enough "environments" that would be seen in each stoichiometry, i.e. the specific bond lengths and atom-atom pairs that would appear in those structures. For this to be the case, I would recommend active learning on an amorphous structure of the material, ideally including a few simple stoichiometries (like 25%, 50%, 75%, as relevant, which can be realized in smaller supercells), in order to obtain an MTP that then can be applied to large-scale structure of varying percentages.

[evansjc]

Hi Anders,

Great! Thank you!