When do I need the se_e3 descriptor?

[youngjae142]

I noticed that there is a new type of descriptor namely se_e3 a three-body descriptor (and I'm very interested).

My question is that,

1. When would it be better to use the new se_e3 descriptor? (or worse when?)
2. Is the se_e3 descriptor better for materials with many types of atomic species? (like ternary compounds?)
   (I'm currently training a ternary compound with se_e2_a descriptors, but it seems not as "good" as the binary case.)

I have been waiting for a report about se_e3 from your group, but I just couldn't wait.
Plz help! 😊

[amcadmus]

e2_a stands for two-body embedding and e3 stands for three body embedding.
In the e2_a descriptor, the three body information is included in the environment matrix but not in the embedding matrix.
e3 introduces three body information in the embedding matrix, thus is more expressive then e2.
In most cases we have tested, the e2_a descriptor is powerful enough, but in some cases like the BCC metal W, one does need e3 to describe the transition barrier of crew dislocations migration.

[youngjae142]

Thanks for your instant answer!
So, would it be the d-orbital's three-body bonding character that requires the three-body embedding matrix?
I know that interpreting the results from such a learning machine is very tricky, but I'm just curious.