Single-model uncertainty quantification in neural network potentials does not consistently outperform model ensembles

Code for performing uncertainty quantification(UQ) for neural network(NN) interatomic potentials using single deterministic NNs and NN ensemble. The software was based on the paper "Single-model uncertainty quantification in neural network potentials does not consistently outperform model ensembles", and implemented by Aik Rui Tan. The code was adapted from the NeuralForceField repo and Atomistic-Adversarial-Attack repo.

The folder contains systems contains script to run training and adversarial attack on the rMD17, ammonia and silica data sets.

The full atomistic data set for:

• rMD17 is available at https://figshare.com/articles/dataset/Revised_MD17_dataset_rMD17_/12672038.
• ammonia is available at https://doi.org/10.24435/materialscloud:2w-6h.
• silica is available at https://doi.org/10.24435/materialscloud:55-sd.

Citing

The reference for the paper is the following:

@misc{tan2023singlemodel,
      title={Single-model uncertainty quantification in neural network potentials does not consistently outperform model ensembles}, 
      author={Aik Rui Tan and Shingo Urata and Samuel Goldman and Johannes C. B. Dietschreit and Rafael Gómez-Bombarelli},
      year={2023},
      eprint={2305.01754},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}