Physics-Informed Machine Learning for Solar Cell Materials Discovery

A comprehensive tutorial and framework for applying physics-informed machine learning to photovoltaic materials discovery using Materials Project data.

Overview

This repository demonstrates how to:

• Extract photovoltaic-relevant properties from Materials Project database
• Construct physics-based descriptors for solar cell materials
• Build interpretable ML models with physical constraints
• Accelerate materials screening while maintaining scientific rigor

Key Philosophy: Rather than treating ML as a black box, we incorporate fundamental photovoltaic physics (band structure, optical absorption, charge transport) into the learning framework.

Quick Start

from src.data_extraction import MaterialsProjectExtractor
from src.descriptors import PhotovoltaicDescriptors
from src.models import PhysicsInformedModel

# Extract materials data
extractor = MaterialsProjectExtractor(api_key="YOUR_API_KEY")
materials = extractor.get_photovoltaic_candidates(n_materials=100)

# Compute physics-based descriptors
descriptor_calc = PhotovoltaicDescriptors()
features = descriptor_calc.compute_all(materials)

# Train physics-informed model
model = PhysicsInformedModel(enforce_bandgap_constraint=True)
model.fit(features, target='efficiency')

Tutorial Notebooks

1. Materials Project Data Extraction (01_materials_project_data_extraction.ipynb)

   • API setup and authentication
   • Querying relevant materials for PV applications
   • Data cleaning and preprocessing

2. Physics Descriptors for Photovoltaics (02_physics_descriptors_for_photovoltaics.ipynb)

   • Electronic band structure descriptors
   • Optical absorption features
   • Charge transport metrics
   • Stability indicators

3. ML Models with Physics Constraints (03_ml_models_with_physics_constraints.ipynb)

   • Incorporating Shockley-Queisser limit
   • Band gap optimization for different architectures
   • Multi-objective optimization (efficiency, stability, cost)

4. Interpretable Predictions (04_interpretable_predictions.ipynb)

   • Feature importance analysis
   • Physical interpretation of ML decisions
   • Uncertainty quantification

🔬 Physics-Informed Features

Electronic Structure

• Band gap (Eg): Optimal range 1.1-1.7 eV for single junction
• Effective masses (m)*: Impacts charge mobility
• Band alignment: Valence/conduction band positions
• Density of states: Near band edges

Optical Properties

• Absorption coefficient: Direct vs indirect transitions
• Spectral matching: Solar spectrum overlap
• Theoretical efficiency: Based on detailed balance

Stability & Synthesis

• Formation energy: Thermodynamic stability
• Decomposition energy: Against competing phases
• Synthesizability score: Likelihood of experimental realization

📊 Example Results

Descriptor Correlation with Efficiency

Physics-Constrained Predictions

Installation

# Clone the repository
git clone https://github.com/NabKh/Physics-Informed-ML-Solar-Cells.git
cd Physics-Informed-ML-Solar-Cells

# Create conda environment
conda env create -f environment.yml
conda activate piml-pv

# Or use pip
pip install -r requirements.txt

Materials Project API Key

1. Register at Materials Project
2. Get your API key from dashboard
3. Set environment variable: export MP_API_KEY="your_key_here"

📖 Key Concepts

Why Physics-Informed ML?

Traditional ML approaches may learn spurious correlations and violate physical laws. Our approach:

1. Constrains predictions to physically reasonable ranges (e.g., Eg > 0)
2. Uses domain knowledge in feature engineering
3. Ensures interpretability through physics-based features
4. Incorporates physical laws (e.g., detailed balance limit)

Shockley-Queisser Limit Integration

def theoretical_efficiency(bandgap, temperature=300):
    """
    Calculate theoretical efficiency limit based on detailed balance.
    Incorporates fundamental thermodynamic constraints.
    """
    # Implementation based on SQ limit
    pass

Educational Goals

This repository serves as:

• Tutorial for students learning ML for materials science
• Template for researchers starting PV materials projects
• Best practices guide for physics-informed ML

📈 Use Cases

1. Rapid Screening

Screen 50-100 materials in minutes instead of months of DFT calculations

2. Design Space Exploration

Identify promising regions in composition-property space

3. Hypothesis Generation

Suggest unconventional materials for experimental validation

4. Teaching Tool

Demonstrate integration of physics and machine learning

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Submit a pull request

📧 Contact

Nabil Khossossi

• Website: sustai-nabil.com
• Email: n.khossossi@differ.nl
• GitHub: @NabKh

📄 License

MIT License - see LICENSE file for details

---

Note: This is an educational and research tool. For production applications, additional validation and testing are required.