Data-efficient and interpretable inverse materials design using a disentangled variational autoencoder

Cheng Zeng, Zulqarnain Khan, Nathan Post. Accepted at AI & Materials

To run the notebooks, some relevant Python packages should be installed, including pyro, shap, pandas, seaborn, numpy, scipy and scikit-learn.

Folder structures

• Data

  • HEA_feature_engineered.csv: experimental high-entropy alloy dataset with engineered features.
  • HEA_top30_comps.csv: experimental high-entropy alloy dataset with 30-element composition features.
  • hyper-parameter-tuning.json: hyperparameter tuning results for scikit-learn neural network models. The key follows the format [alpha]-[hls]-[lr] where alpha, hls and lr refer to L2 regularization strength, hidden layer structure and learning rates, respectively.
  • test_data_reconstruction_analysis.csv: alloy reconstruction analysis results for all 138 test data points.
  • look_up_dict.pkl: look-up dictionary saved in pickle format to calculate engineered features for an arbitray alloy.
  • mixing_enthalpy_dict.pkl: mixing enthalpy look-up dictionary in pickle format
  • labelled_hea.pk, unlabelled_hea.pk, validation_hea.pk, test_hea.pk: pickle files for different types of data for training the model used for inference.

• models

  • ssvae.model: semi-supervised variational autoencoder model with the highest test accuracy 0.877. This model is used for all inference results shown in the manuscript.

• notebooks

  • 1_SSVAE_Model_training.ipynb: for training the ssvae model.
  • 2_SSVAE_model_inference.ipynb: for predictions and generating most figures present in the manuscript.
  • 3_SSVAE_SHAP_analysis.ipynb: shapley analysis using a kernel-based method. Including the overall and individual feature importance results.
  • 4_SSVAE_Alloy_reconstruction.ipynb: Alloy reconstruction study and summary for 138 test data points.
  • 5_SSVAE_Interpolation_study.ipynb: Interpolation study for the generative model focusing on a small latent region corresponding to refractory alloys.

• utils

  • custom_mlp.py: basic classes for building the SSVAE model.
  • featurization.py: some utility functions to calculate the composition feature vector and engineered feature vectors for a given chemical formula.