Implementing the Single Neural Network in N2P2

[omorrison1]

To my knowledge N2P2 is purely designed for the Behler Parinello Neural Networks (BPNNs).

An atomic neural network is trained for each element.

Since this has a scaling problem (more elements dramatically increases the number of descriptors), I was wondering if it possible to train a "single neural network" model (S-NN) through N2P2. If not, would it be possible to add this model as an option.

In contrast to BPNN, the S-NN model assigns the same atomic neural network to every element, with the atomic number of the centre atom used as a descriptor.

Here is a paper in case you are curious: https://pubs.acs.org/doi/pdf/10.1021/acs.jpcc.0c04225

Thanks, Oliver Morrison

[singraber]

Hello!

That's certainly an interesting approach and it would be great to have that feature. Also, some parts of the n2p2 code base are modular enough to reuse their functionality for this purpose. However, I simply don't have the time to do this for you. If you were willing to lead the development of this feature, I'd be happy to assist whenever I can.

Best, Andreas