search

qzhu2017 / PyXtal_ml

a Python3 library for ML modeling materials properties
MIT License
11 stars 1 forks source link

[Concern] Crystal graph #40

Closed yanxon closed 5 years ago

yanxon commented 5 years ago

@qzhu2017

https://github.com/qzhu2017/PyXtal_ml/blob/b11b3405996b4e587ab014e9727239f59440a042/pyxtal_ml/descriptors/crystal_graph_QZ.py#L68

I just read your comments in the code.

I think the GaussianDistance is the secret sauce of crystal graph. Without GaussianDistance, nbr_fea is just distances between two atoms without the mapping.

yanxon commented 5 years ago

@qzhu2017

Also, let me work on the code.

Thank you,

Howard

qzhu2017 commented 5 years ago

Just to clarify, the learning of images and graphs should be very different.

@yanxon The way which we understand CNN is used for image processing. Graphs is very different from 2D images, as we care more about the nodes, edges, and weight. These things can create a lot of discontinuity if we treat them as a 2D image. So a lot of methods have been developed by the community, the real name is called Graph Convolutional Networks (GCN). The crystal graph should belong to GCN problem instead of CNN.

qzhu2017 commented 5 years ago

http://tkipf.github.io/graph-convolutional-networks/

yanxon commented 5 years ago

I see. Thank you for the info. I will work on this and the gaussian descriptor. Hopefully, I can get them done by Monday.