title	emoji	colorFrom	colorTo	sdk	sdk_version	app_file	pinned	license	short_description
TransPolymer - Polymer Property Predictor	🧪	blue	purple	streamlit	1.28.0	app.py	false	mit	AI-powered polymer property predictions using Transformer architecture

🧪 TransPolymer: Polymer Property Predictor

TransPolymer is a Transformer-based language model for polymer property predictions, deployed as an interactive web application on Hugging Face Spaces.

🚀 Features

• 10+ Property Predictions: Support for various polymer properties including:

  • Polymer Electrolyte Conductivity (PE_I, PE_II)
  • Organic Photovoltaic Efficiency (OPV)
  • Band Gap (Egc, Egb)
  • Electron Affinity (Eea)
  • Ionization Energy (Ei)
  • Crystallization Tendency (Xc)
  • Dielectric Constant (EPS)
  • Refractive Index (Nc)

• SMILES Input: Easy polymer structure input using SMILES notation

• Batch Predictions: Upload CSV files for bulk predictions

• Interactive Interface: User-friendly Streamlit interface

🔬 Model Architecture

• Base Model: 6-layer RoBERTa architecture with 768 hidden dimensions
• Training: Pretrained on ~5 million polymer SMILES sequences
• Tokenization: Custom chemical-aware tokenization for polymers
• Framework: PyTorch + Transformers

📖 How to Use

1. Single Prediction:

   • Enter a polymer SMILES string (e.g., CC(C)(C)OC(=O)NC1=CC=CC=C1)
   • Select the property you want to predict
   • Click "Predict Property"

2. Batch Prediction:

   • Upload a CSV file with a 'smiles' column
   • Select the property type
   • Download results with predictions

⚠️ Important Notes

• Demo Version: This Space runs with random weights for demonstration purposes
• Research Tool: Results should be validated experimentally
• Educational Use: Intended for research and educational applications

📚 Citation

If you use TransPolymer in your research, please cite:

@article{xu2023transpolymer,
  title={TransPolymer: a Transformer-based language model for polymer property predictions},
  author={Xu, Changwen and Wang, Yuyang and Barati Farimani, Amir},
  journal={npj Computational Materials},
  volume={9},
  number={1},
  pages={64},
  year={2023},
  publisher={Nature Publishing Group UK London}
}

🔗 Links

• 📄 Paper
• 💻 GitHub Repository
• 🤗 Hugging Face Space

🛠️ Technical Details

The application uses:

• Streamlit for the web interface
• PyTorch for model inference
• Transformers library for model architecture
• Custom tokenization for chemical SMILES processing

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Developed by researchers at Carnegie Mellon University