Questions about SOAP descriptors

[rajathsalegame]

Hi, I had a few questions about the SOAP implementation within dscribe that I was hoping someone could answer:

1. What are some empirical ways to make sure that the SOAP outputs I am seeing are sensible and make sense (either visual or mathematical?
2. What are some strategies to decide on the best hyperparameters for SOAP generation (setting r_cut, n_max, l_max)?
3. Just making sure, but if I specify the positions to calculate as [1,4,10] for a given molecule, the SOAP output will be in the order [SOAP_VECTOR[1], SOAP_VECTOR[4], SOAP_VECTOR[10]], correct?

Thanks!

[lauri-codes]

Hi @rajathsalegame.

1. With SOAP this is quite hard as the features are so abstract. But here are a few tips that come to mind:
   • You can plot the flattened features and color/label different parts based on the ranges that you can get out of soap.get_location(("H", "O"))
   • You can calculate the dot product of normalized SOAP feature vectors to get a rough idea of the similarity of different outputs. The values will be between [0, 1], where 0 means completely different, 1 is completely identical.
2. In a supervised machine learning application, you would include these as hyperparameters to your cross-validation. This is a very standard way to select good ones. For unsupervised learning or other uses, the selection becomes much harder. The pragmatic way would be to test a few settings or do a literature search for values others have used.
3. You are correct.

[rajathsalegame]

Thank you!