Data and models for: Entropy based active learning of graph neural networks for materials properties

Re-creating the Paper

To recreate the experiments performed in the paper you will first need to clone the code in the GitHub repository to your local machine: https://github.com/keeeto/gp-net

Next, set up the right python environment - there is a gp-megnet.yml file in the repository that can be used to create the right environment using conda.

Then take the paper-experiments.tgz from this repository in the top directory of the repository that you cloned from GitHub.

From the directory experiments you can then run the the experiments with the following commands:

python gp-net.py -data formation_energy_per_atom_data.pkl -ndims 1 -q 3000 -frac 0.4 -epochs 0 -prev -cycle 1 3000 -maxiters 100 -stop 0.99 -include

python gp-net-skip-activation.py -data formation_energy_per_atom_data.pkl -ndims 1 -q 3000 -frac 0.4 -epochs 0 -prev -cycle 1 3000 -maxiters 0 -stop 0.99 -include -samp random -amp 0.4281 -length 2.3997

Plots

Details on the data analysis and plotting are included in the .ipynb files.

Details

These are the data and ML models needed to recreate the results in: Entropy based active learning of graph neural networks for materials properties.

* The hdf5 files contain the saved parameters for the trained Megnet models used in this study.

* gp_mae_{random/entropy}.npy contain the results of the active learning procedure with either entropy based or random sampling.

* gp_mean and gp_stdev contain the mean and standard deviation of the fully trained model on the test set data.

* yvals contains the true lablel for the same data as in gp_mean and gp_stedev.

* samp_indices contains the sample indices for  yvals and can be used to work out which material corresponds to which value.

* formation_energy_per_atom.pkl is the training data from the Materials Project used in this study.

* latent_test.npy contains the latent space vectors in 2D.

* yvals_latent,npy contains the formation energies corresponding to the the latent space points in latent_test.npy

The data in this folder can be plotted using notebooks available in the publication_notebooks folder of the github repo for the active learning code: https://github.com/keeeto/gp-net