Charges in LODE (Question)

[atulcthakur]

Hi,

Thanks for a great package. I have been reading the documentation and API to understand more about torch-pme. I am looking at the LODE example and it seems to be needing atomic charges as weights on the grids. As far as I understand, the original LODE does not depend on these atomic charges.
I wanted some clarification about why torch-pme needs atomic charges, when featomic can compute LODE without it (unless I am totally misunderstanding LODE construction)?
Also, what would be a recommended way for computing LODE? Featomic or torch-pme?

[ceriottm]

As you can see, the LODE example is kind of a demo more than a full-fledged implementation. If you read the preprint accompanying the library, our impression is that LODE can be simplified a great deal without much accuracy loss.

If you need a full-fledged LODE implementation consistent with the original version, you better stick with featomic for the moment, but I'm planning to "promote" the example to some kind of rudimentary calculator, that should be much faster on GPUs, and for very large boxes (as well as much more flexible). In all event, if you want to replicate the per-element density of LODE, you need to simply use a one-hot encoding of the atom types (have multiple channels, and assign charges of one for one of the elements in each channel).