Data for: Learning a reactive potential for silica-water through uncertainty attribution

This is the dataset for "Learning a reactive potential for silica-water through uncertainty attribution", by S.Roy, J.Dürholt, T.Asche, F.Zipoli, R.Gómez-Bombarelli. The repository contains neural network interatomic potentials (NNIP) ensemble used for running MD simulations in the paper. This potentials are torch models that can loaded with torch.load. The potential can be used in https://github.com/learningmatter-mit/NeuralForceField for replicating calculations in the paper. The details on how to use it can be found in the github repository. The contents of the compressed folder Painn_ensemble are:

• log_human_read.csv: contains all training details of three PaiNN models.
• model_0: folder for model with id 0 same for model_1 and model_2

Each model_0:2 folder includes:

• log.csv: contains all training details of the PaiNN model.
• best_model: contains the best PaiNN model with lowest error that can be read using torch.load and is used in our study
• checkpoints: folder that contains PaiNN model checkpoints for final three epochs. They are not used and stored only for emergency uses.

Our NNIP ensemble consists of three PaiNN models which are the best_model in their model_id folders. Out of which model_2/best_model is our best PaiNN model which we use for single model simulations. The potentials output energy and forces and also stresses if required. The units for energy is given in kcal/mol, and the units for forces are given in kcal/mol Å and units for stresses are kcal/mol Å^3. Positions and lattice parameters are given in Ångstrom.