Materials Project Database

In this example, we will create and train a neural network on the property formation_energy_per_atom based on the architecture as introduced in this paper, using the Materials Project database.

1. Set up the Dataset

a. Set up required dependencies

Create an API key with Materials Project to download the training dataset for this example.

Python packages (primarily pymatgen) are also required as dependencies. The easiest way to install these in a new environment, is using Conda, by running the following commands.

conda create -name agn_example1
conda activate agn_example1
conda install -c conda-forge pymatgen=2022.0.4

Note: Alternatively, to install dependencies in an existing environment, you can skip the first step, activate your environment, and directly install pymatgen=2022.0.4.

b. Download the Data

To download the structures, simply run

python download_data.py "YOUR_API_KEY"
where, "YOUR_API_KEY" is replaced with your actual API key obtained from the Materials Project database.

This downloads the data required to train the models into the directory example1/data.

2. Train the Network

Now that the dataset has been set up, run the formation_energy.jl file in your Julia environment and play around with the training function!

Note

• By default, the property chosen for training is formation energy per atom. If you want to train a different property, replace it with its pymatgen string in the appropriate line. This change must be reflected in both download.py and example.jl.

• There may be some CIF parsing warnings from pymatgen that show up, but these shouldn't affect things and can be safely ignored.