Reflecto: Deep Learning for X-ray Reflectivity Analysis

Automated extraction of thin film parameters (thickness, roughness, electron density) from X-ray Reflectivity measurements using deep learning.

---

🎯 Overview

The Problem:

• Traditional XRR analysis relies on manual fitting
• Time-consuming and requires expert knowledge
• Sensitive to initial parameter guesses

Our Solution:

• Physics-based simulation for training data generation
• End-to-end learning with CNN
• Real-time analysis (seconds)

---

Preperation before use

Windows

# Install uv
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex

# Clone Repository
git clone https://github.com/SJB7777/reflecto

# Make virtual environment
uv venv --python 3.13
.venv/scripts/activate

# Install packages
uv sync
uv run python -m ipykernel install --user --name=my-env --display-name="reflecto"
uv pip install -e .

---

🚀 Quick Start

# Run entire pipeline (data generation → training → evaluation)
python runs/exp05_1layer_mask/main.py

That's it! The script automatically handles data generation, model training, and evaluation.

---

📁 Project Structure

runs/
├── exp01_quan_class/      # Classification approach for multi-layer
├── exp02_one_layer/       # Regression for single layer (RefNX)
├── exp03_physics_fused/   # Physics-guided learning (experimental)
├── exp04_one_genx/        # Regression for single layer (GenX)
├── exp05_1layer_mask/
└── exp06_1lay_6out/

Each experiment contains:

• config.py - Experiment settings
• main.py - Full pipeline runner
• model.py - Neural network architecture
• evaluate.py - Performance analysis

---

⚙️ Configuration

Edit config.py to customize:

CONFIG = {
    "param_ranges": {
        "thickness": (20.0, 200.0),  # Å
        "roughness": (0.0, 10.0),    # Å
        "sld": (0.0, 140.0),         # 1e-6 Å^-2
    },
    "simulation": {
        "n_samples": 1_000_000,
        "q_points": 200,
    },
    "training": {
        "batch_size": 128,
        "epochs": 50,
        "lr": 0.001,
    }
}

---

🔬 Experiments Comparison (Keep develping)

Experiment	Target	Approach	Best For
exp01	Multi-layer	Classification	Fast screening
exp02/04	Single layer	Regression	High accuracy
exp03	Multi-layer	Physics-fused	Interpretability

Recommendation: Start with exp02 or exp04 for single-layer analysis.

---

📊 Performance

Typical results on simulated data (single layer):

• Thickness: MAE ~2Å
• Roughness: MAE ~1Å
• SLD: MAE ~0.15 (×10⁻⁶ Å⁻²)

---

🔧 Key Features

Simulation:

• RefNX/GenX-based XRR generation
• Realistic noise models (Poisson + background)
• Physical constraints enforcement

Training:

• Automatic checkpoint management
• Mixed precision support
• Resume from interruption

Evaluation:

• Comprehensive metrics (MAE, RMSE)
• Visualization tools (parity plots, error distribution)
• Real experimental data support