Deep Adaptive Regressive Weighted Intelligent Network (DARWIN)

This Repository implements the DARWIN approach for interpretable discovery of materials. In the process, the package provides following functionalities:

• Enables search for promising materials with specified target properties using evolutionary algorithm
• Enables interpretability analysis on the predictions of the machine learning model and evolutionary algorithm
• Provides trained and ready-to-use machine learning models for energy-above-hull, bandgaps and direct-indirect nature of the bandgap
• Transfer learning of pre-trained models
• The following paper describes the details of the DARWIN framework: DARWIN

Table of Contents:

• Requirements
• Usage
  • Using existing prediction
  • Using EA with pretrained models
  • Interfacing DARWIN with custom models
  • Transfer learning
• Data
• Authors
• License
• How to cite

Requirements:

DARWIN shares all the dependencies with MatDeepLearn that was forked and used for performing hyper-parameter optimizations for different possible neural networks. Therefore, please follow the following steps to use DARWIN:

• Clone this repository
• Install all the MatDeepLearn requirements
• We provide a fork of MatDeepLearn to enable predictions on-the-fly for smaller sets of materials
• Once all the requirements are installed, you should add DARWIN to your system path as:

import sys
sys.path.append('location-of-DARWIN')

Data:

• All the data used in this study was obtained from OQMD (for pre-training the model), Materials Project (for fine-tuning on energy-above-hull and indirect-direct nature of bandgap) and SUNMAT (for fine-tuning on HSE06 bandgaps)

Usage:

Using EA with models of this study:

• You can use our pre-trained models to search for desirable targets and generate insights from the run
• Please use the python main.py --output interpretability_groups.pkl --generational generational_data.pkl for performing the analysis.
• You can specify the set or subset of properties that are of interest in the main function and their corresponding ranges.
• The generated candidates will be stored at the output filename provided above.
• All the candidates, their corresponding properties and fitness (we use the term "badness" which means negative of the fitness) values are stored in the generational filename.
• Use interpretability.ipynb for running and visualizing the analysis.

Using existing predictions:

• It is possible to use DARWIN with one's pre-existing predictions.
• The predictions should be provided in two columns: control and treatment.
• You can run the DARWIN analysis using the notebook provided in the repository.
• Methods available as of now for analysis are: spearman (Kruskal-Wallis test and spearman correlation) and Random Forest permutation imporatance.
• The notebook also provides a clean way to represent the findings in a histogram formatted with latex.

Interfacing DARWIN with custom models:

• Yet another way to run DARWIN is on models not provided by us
• Users can choose the modules they want to use: EA or unsupervised analysis or both

Transfer learning:

• You can find the support to perform transfer learning using the TL module.
• We provide checkpoints for two pre-trained models that have been trained to predict OQMD formation energies from the relaxed structures.

Authors

This software was written by Hitarth Choubisa.

License

DARWIN is released under the MIT License.

How to cite:

Please cite the following work if you use DARWIN:

@article{Choubisa2023,
author={Choubisa, Hitarth
and Todorovi{\'{c}}, Petar
and Pina, Joao M.
and Parmar, Darshan H.
and Li, Ziliang
and Voznyy, Oleksandr
and Tamblyn, Isaac
and Sargent, Edward H.},
title={Interpretable discovery of semiconductors with machine learning},
journal={npj Computational Materials},
year={2023},
month={Jun},
day={29},
volume={9},
number={1},
pages={117},
issn={2057-3960},
doi={10.1038/s41524-023-01066-9},
url={https://doi.org/10.1038/s41524-023-01066-9}
}