Question about result recurrence

[loilisxka]

Hello, I'm trying to reproduce the Ag3+/- results shown in the spookyNet paper, but there seems to be a problem with the model training.

The training code refers to issue1, and the code file is in the "code.zip". The energy RMSE obtained by training is 26.51meV/atom. The parameters are set according to the paper. I don't know why the model was not trained correctly.
code.zip

I use energy and force for error backpropagation. And modifications were made in the energy_and_forces function of spookynet.py. The modified code is:

grad = torch.autograd.grad(
                [torch.sum(energy)], [R], create_graph=create_graph, retain_graph=True
            )[0]

[ShawnBox]

Hi, I think I meet the same problem as you. So I'm wondering if you have managed to resolve it? If so, could you please share how you did it?

[loilisxka]

This problem has not been solved yet, so I gave up using this model.