1) Embed asymmetric unit molecule to single node in an equivariant fashion. (optionally in-future via pretrained encoder)
2) Generate crystal graph either by explicit symmetry patterning or some other clever means.
3) Evaluate a second invariant GNN on the molecule-wise graph, to produce outputs.
Should be trainable and if it converges hopefully will be significantly faster than current methods, and open opportunities for new kinds of crystal embeddings.
1) Embed asymmetric unit molecule to single node in an equivariant fashion. (optionally in-future via pretrained encoder)
2) Generate crystal graph either by explicit symmetry patterning or some other clever means.
3) Evaluate a second invariant GNN on the molecule-wise graph, to produce outputs.
Should be trainable and if it converges hopefully will be significantly faster than current methods, and open opportunities for new kinds of crystal embeddings.