jeherr / elpasolite-formation-energy-predictor

Elpasolite Formation Energy Predictor

Neural network model to predict the formation energy for elpasolites of the crystal formula ABC₂D₆.

Prerequisites

Required packages

numpy
tensorflow

Getting Started

To download this project run the following command

git clone https://github.com/jeherr/Elpasolite-Formation-Energy-Predictor.git

ElpasEM_Thu_Mar_21_13.16.26_2019 contains an already trained network model which can be used to make predictions for any combination of main group elements up to Bismuth at any of the crystal lattice positions. To predict the formation energy (in eV) for any elpasolite run

python evaluate.py A B C D

where A, B, C, and D are integers corresponding to the desired atomic number at each crystal lattice position. For example

python evaluate.py 13 11 19 9

will print the prediction for the prototypical elpasolite AlNaK₂F₆.

Training your own model

To train your own model run

python train.py train_set.pkl

which uses the same hyperparameters laid out in the paper and stops after 1000 epochs. The model is evaluated on the validation set after every 5 epochs and prints out a random sample of 10 true formation energies and the corresponding predictions, along with mean absolute errors, mean signed errors, and root mean square errors over the validation set. Only saves a new checkpoint if the evaluation loss is lower than the last saved checkpoint. After training finishes, the best checkpoint is reloaded and the errors are evaluated over the test set. Trained model create a new directory with the data and time the model was started. To use your own model with evaluate.py, replace the default model directory with the directory for your newly trained model in the following line.

model = network_model.NNModel(name="ElpasEM_Thu_Mar_21_13.16.26_2019")

Citing this work

A publication for citation available at https://doi.org/10.1063/1.5108803.

Acknowledgments

Thanks to the work linked below for the data set of elpasolites and formation energies used here to train the model. https://doi.org/10.1103/PhysRevLett.117.135502