Bandgap_Model: Bandgap Prediction for Solid-State Materials

Bandgapg_Model is a Patolli-based framework for predicting bandgaps in crystalline (solid-state) materials and for exploring AI-driven approaches to accelerate materials discovery.
It combines classical machine learning and deep learning methods with extensive feature engineering and analysis, providing a versatile platform for predictive experimentation in materials science.

---

Features

• General Exploration and Experimentation Notebook: From descriptor generation and analysis to deep learning architecture training.
• Feature Engineering & Dataset Preparation: Generate descriptors and prepare unified datasets for model training.
• Concatenated Descriptor Exploration: Combine CNN-extracted features with Patolli descriptors for ML and DL modeling.
• FGSM Analysis: Evaluate model behaviour to adversarial perturbations.
• Feature Abstraction Analysis: Examine intermediate representations from the penultimate layer of top-performing models.
• Bandgap Simulator: Predict the bandgap of user-defined compounds (perovskite) using trained models.

---

Repository Contents

BgModel/
├── BandgapModel.ipynb                  # Main exploration (Patolli descriptors)
├── Descriptor_Concatenation.ipynb      # Building concatenated CNN + Patolli datasets
├── BandgapModel_Concatenation.ipynb    # CNN + Patolli combined features
├── BandgapModel_FGSM.ipynb             # Robustness analysis with FGSM
├── BandgapModel_FeatureAbstraction.ipynb # Penultimate layer feature analysis
├── ChemicalSpace_Dataset_Generation.ipynb # Materials API dataset building
├── BandgapSimulator_Perovskites.ipynb # Custom bandgap simulator
│    
│
├── data/      # Training and test datasets (limited due to file size)
├── models/    # Trained models (limited due to file size)
├── README.md
├── .gitignore
└── requirements.txt

---

Dependencies

• Python 3.12.4
• TensorFlow 2.17
• keras
• scikit-learn
• pandas
• numpy
• matplotlib
• XGBoost
• pydot
• itertools

Use pip install -r requirements.txt to install all dependencies.

---

Usage

All notebooks are interactive and can be run in Jupyter or Colab.

My workflow:

1. Explore and analyze Patolli-generated descriptors exclusively, including model architectures and training, in DL_Model_Exploration.ipynb.
2. Generate concatenated datasets that include CNN-extracted features from X-ray diffractogram models in FeatureEngineering_and_DatasetPreparation.ipynb.
3. Investigate the combined descriptor spaces in Concatenated_Descriptor_Analysis.ipynb.
4. Adversarial attack implementation (FGSM) in FGSM_Analysis.ipynb.
5. Analyze feature abstractions from top-performing models using FeatureAbstraction_PenultimateLayer.ipynb.
6. Simulate bandgap energy values for custom compounds in Bandgap_Simulator.ipynb.

---

Datasets

The data/ folder contains training and test sets.

Note: Trained model files and datasets are provided selectively due to their large size. Users can train models using the notebooks.

---

Citation

This work is based upon:

J. I. Gómez-Peralta, X. Bokhimi. Journal of Solid State Chemistry, Vol. 285 (2020) 121253.

GitHub repositories: 1. Patolli 2. Perovskite_Simulator

This project was developed as part of a research internship at LANNBIO-CINVESTAV.

---

Notes

• Users can modify descriptors or model architectures to explore new material systems.

---