BPNNP-pytorch
The code is used for constructing interatomic force field within the framework of Behler-Parrinello Neural Network Potential (BPNNP). If you are enough with n2p2 or RuNNer's slow training, you can take a look. By using of GPU computing and pytorch's automatic differentiation, the training can be accelerated for at least 20 times.
Just directly use the jupyter notebook for training now.
Things already done
- Dataloader which accepts ASE database file as the input.
- 4 symmetry functions as defined in Atom-centered symmetry functions for constructing high-dimensional neural network potentials.
- Fast neighbor list algorithm using GPU.
- Model training with simple full connected neural networks (FCNN).
Future updates
- Use better organized inputs for training the model.
- ASE machine learning calculator.
- Add Bayesian layers in the model.
- Batch active learning workflow.