Project 2 - Machine Learning

Crystal Structure Descriptor for Machine Learning

Stores code for our experiments with a simple crystal structure descriptor for machine learning on binary materials

Requirements

Use the package manager pip to install any required packages for the project. You will need LightGBM, TensorFlow, sklearn, pandas, and jupyter. Note that future iterations of the project will require PyTorch too. Please look up their respectives websites to install them.

Running the project

There are three different machine learning models implemented for each of the two tasks. Please refer to the Project Structure for the following section.

Folder 'conduction_models' includes 4-layer neural network using TensorFlow in '4Layer_TF_binary.ipynb', LightGBM method in 'LGBM_Boosted_Trees.ipynb', and logistic regression in 'LogisticRegression.ipynb'. These models are used for binary classification of materials as conductors/insulators.

The predictions on the formation energies of the materials are obtained using 4-layer neural network in '4Layer_net_TF.ipynb', LightGBM method in 'LGBM_Boosted_Trees.ipynb', and linear regression in 'Linear Regression.ipynb' from the 'energy_models' folder.

All the models run using the data processed by our crystal-structure descriptor from file 'Descriptor.csv' at folder 'data/descriptor'. This file can be re-generated by running 'descriptor_run.ipynb' file from 'data_processing' folder. The initial data is provided in files 'binary_cell_df.pkl' and 'binary_species_list.pkl' in the folder 'data/pkls'.

The original data was fetched from the Materials Project database (https://materialsproject.org/). We fetch the following properties for all binary compounds ("nelements":2): 'energy_per_atom', 'band_gap', 'initial_structure'.

Project Structure

Model files:

├── energy_models
│   ├── 4Layer_net_TF.ipynb
│   ├── LGBM_Boosted_Trees.ipynb
│   └── Linear Regression.ipynb
└── conduction_models
    ├── 4Layer_TF_binary.ipynb
    ├── LGBM_Boosted_Trees.ipynb
    └── LogisticRegression.ipynb

Data files:

└── data
    ├── descriptor
    |   └── DescriptorData.csv
    └── pkls
        ├── binary_cell_df.pkl
        └── binary_species_list.pkl

Extras(Plots and others):

├── README.md
├── energy_models
│   └── Accuracy.pdf
└── conduction_models
    ├── Plotting.ipynb
    └── ROC_binary.pdf

Utility files:

├── data_processing
|   ├── descriptor.py
|   ├── descriptorExample.ipynb
|   └── descriptor_run.ipynb
└── helpers
    ├── data_utils.py
    └── picklers.py