Training and prediction of structures without 3-body interactions

[JiQi535]

I have verified that the current modifications can ensure reliable prediction and training of any number of structures without 3-body interactions. By contrast, the previous solution in TF version (see related discussions) has to make a static graph with 3-body interactions beforehand. That solution also cannot ensure reliable training when there are a whole batch of structures without 3-body interactions in the training set. This is actually a possible case when users manually add isolated structures into training set.

To be exact, the edited code is tested to perform a few epochs of training for 3 different sets, (1) 50 structures with isolated atoms, (2) 50 structures with only 2-body interactions, and (3) 407 Si-O structures in the Materials Project (same as the example of M3GNet training with matgl). For prediction, the edited code is expected to work for any structures with or without 2- or 3-body interactions. These edits are expected to bring ignorable effect to speed and memory consumption of training or prediction.

@shyuep @kenko911 Would you kindly have a look and merge this PR if no further questions? Thanks!

fixed #78
closed #85

[JiQi535]

@kenko911 Thanks for the helpful and inspiring discussions. I have edited this PR to combine our ideas. Now, no artificial bond is created. Instead, non-bonded nodes are used to better represent isolated atoms. This current solution is expected to work for training and prediction of any structures, including any combination of isolated atom, two-body and three-body interaction regions. I will continue to review and see if this version will be able to enable bias properly. Thanks.

[kenko911]

@JiQi535 Thanks a lot for your effort!!