Accelerating Search for Superconductors using Machine Learning

Description

This repository contains code used for my project: Accelerating Search for Superconductors using Machine Learning.

Requirements

• Python 3.10+ is required. You can check your Python version using:

  python3 --version
  

Usage

• Clone this repository to your local machine:

  git clone https://github.com/adigasuhas/Accelerating-Search-for-Superconductors-using-Machine-Learning.git

Using pre-trained models

• Download the trained models from Google Drive: Click Here

• Ensure to rename the unzipped folder as Machine_Learning_Models.

• Move the Machine_Learning_Models folder to the cloned repository:

  mv Machine_Learning_Models Accelerating-Search-for-Superconductors-using-Machine-Learning/
  

(OR)

Training model with Quantum Structure Diagram based descriptors

To reproduce the results reported in the manuscript, use the trainer function provided in the Training folder.

A convenient bash wrapper script Trainer.sh is also included to execute training with minimal commands.

If the script does not have execute permission, grant it and run it as follows:

cd Accelerating-Search-for-Superconductors-using-Machine-Learning
chmod +700 Trainer.sh
./Trainer.sh

• Command Line Options:

  • --save - Saves the trained classification and regression models as joblib files in the Machine_Learning_Models directory
  • --model30 - Trains a random forest regression model using 30 QSD-inspired descriptors used in the manuscript.
  • --model5 - Trains a random forest regression model using 5 key features identified through SHAP analysis.

• The full training source code can be found in Training/Trainer.py

Note: If you retrain the models yourself, a minor numerical variations in predictions may occur. This happens because the dataset is randomly split into training and test sets each time, slightly changing the fitted parameters, even though all hyperparameters remain same.

Preparing the Input File

• Navigate to the code directory where the $T_c$ prediction scripts are located:

  cd Accelerating-Search-for-Superconductors-using-Machine-Learning/Temp_Predictor/
  

• Open the file named Material_prediction.csv.

• The following columns can be handled as described:

  • Material-ID (Optional): This is for user reference and identification. It is not auto-filled by the code, so you may enter any identifier (e.g., Unique Identification number, Sample name).
  • Chemical_Formula: Add the materials (chemical compositions) for which you wish to predict the critical temperature ($T_c$).
  • Temp_critical(Optional): If known, you may enter the experimental $T_c$ here for comparison. Otherwise, you may leave it blank.

• Follow the formatting rules illustrated in the reference image provided in the repository to ensure your chemical composition input is valid.

  

Compound Similarity Check

• After entering the Chemical_Formula into Material_prediction.csv, you can check if the compound is present in the SuperCon-MTG dataset by running the following command. A warning will be displayed, and if the compound is found, its corresponding material ID and critical temperature will be shown.

cd Temp_Predictor
python3 Compound_Matcher.py

Generating Descriptors and Critical Temperature Prediction

• Once you have validated that the compound is not present in the SuperCon-MTG dataset, you can proceed to generate descriptors for $T_c$ prediction. To do this, simply run the Predict.sh script, which will generate the descriptors and predict $T_c$ using both the 30-feature and 5-feature models.

  If the script does not have execute permissions, you can grant them using the following command (skip if already granted):

  chmod +700 Predict.sh
  ./Predict.sh
  
  

• Upon successful execution, a Material_prediction_results.csv file will be generated inside the Temp_Predictor folder. This file will contain the following columns:

  • Material-ID
  • Chemical_Formula
  • Temp_critical
  • Predicted_class
  • Predicted_Temp_critical_30_Features
  • Predicted_Temp_critical_5_Features

References

This work has been made possible due to the insights and datasets provided by the following sources, which have served as key inspirations:

1. Original Dataset:

   • Center for Basic Research on Materials. MDR SuperCon Datasheet Ver.240322. National Institute for Materials Science, 2024. DOI: 10.48505/NIMS.4487

2. Descriptors:

   • Rabe, K. M., Phillips, J. C., Villars, P., & Brown, I. D. (1992). Global multinary structural chemistry of stable quasicrystals, high-$(T_c)$ ferroelectrics, and high-$(T_c)$ superconductors. Phys. Rev. B, 45(14), 7650–7676. DOI: 10.1103/PhysRevB.45.7650

3. Composition Generation:

   • Davies, D., Butler, K., Jackson, A., Skelton, J., Morita, K., & Walsh, A. (2019). SMACT: Semiconducting Materials by Analogy and Chemical Theory. Journal of Open Source Software, 4(38), 1361. DOI: 10.21105/joss.01361

4. Machine Learning-Based Work:

   • Stanev, V., Oses, C., Kusne, A. G., Rodriguez, E., Paglione, J., Curtarolo, S., & Takeuchi, I. (2018). Machine learning modeling of superconducting critical temperature. npj Computational Materials, 4(1). DOI: 10.1038/s41524-018-0085-8

Our Work

If you'd like to cite our work.

@misc{Adiga2025,
      title={Accelerating the Search for Superconductors Using Machine Learning}, 
      author={Suhas Adiga and Umesh V. Waghmare},
      year={2025},
      eprint={2505.11964},
      archivePrefix={arXiv},
      primaryClass={cond-mat.supr-con},
      url={https://arxiv.org/abs/2505.11964}, 
}