How and why train the e3nn model

[WJiangH]

Dear Authors,

I have a question regarding the training of the GNN to represent the graph's latent embeddings. Based on my understanding, the weights in the e3nn model utilized in your research are randomly initialized and remain untrained. I confirmed this by initializing another network in your example case for ternary systems. Interestingly, the results, such as the number of distinct 1CP and the Dissimilarity between two embedded graphs, remained unchanged despite the differing weights in the e3nn, which resulted in varied embedding values. I thought that changes in embedding values might influence the PCA results and thereby affect the dissimilarity evaluations, but it appears I might be missing some insights.

Also, I noted that in your "ml.py" file, the model was indeed trained and the network's weights were loaded in the example case. Could you please clarify how you prepared the dataset and labels for this training, and the purpose behind it?

Best regards,
JJ

[WJiangH]

I apologize for the oversight in my previous comment. Upon retesting some of the codes, I observed that the dissimilarity does indeed vary with different e3nn models when random weights are used. I initially overlooked that the "dataset.py" file preloads the dataset.

Could you please confirm if the ModelTrainer class is intended for use with the AutoEncoder? I noticed that there is an option to use an Autoencoder for embedding the graph in your code, which led me to this assumption.

Thank you for your assistance.

Best regards,
JJ

[killiansheriff]

Hi JJ,

You are correct, the weights are randomly initialized, the network remain untrained, and this has little to none influence on the identification of the 1CP.

The random initialization does has some influence on the embedding space, and hence of the dissimilarity. However, if you design a robust framework to extract information from the correlation functions, then your physical quantity should be, to some extent, independent of the random initialization. I address that point in SI section 7 of https://www.nature.com/articles/s41524-024-01393-5.

Hope this helps, don't hesitate if you have any other questions!

Best,
Killian