Skip to content

PULL_REQUEST.md

Latest commit

 

History

History
90 lines (70 loc) · 3.29 KB

PULL_REQUEST.md

File metadata and controls

90 lines (70 loc) · 3.29 KB
Raw

Powder Diffraction Simulation with PyTorch Acceleration

Summary

This PR introduces a major overhaul of xrd_simulator to support powder diffraction simulation with GPU-accelerated rendering via PyTorch. The refactor enables realistic simulation of polycrystalline samples across all grain size regimes—from nanocrystals to large single crystals—with physically accurate peak broadening models.

Key Features

🚀 PyTorch Backend

  • Complete migration from NumPy to PyTorch for all core computations
  • Automatic CPU/GPU device selection via new xrd_simulator.cuda module
  • Memory-efficient batch processing for millions of grains

🔬 Multi-Scale Rendering Methods

New rendering system with three physically-motivated methods:

Method Grain Size Description
nano < 0.1³ µm³ Airy disk patterns with Scherrer broadening for nanocrystals
micro 0.1³ - pixel³ µm³ Fast Gaussian profiles for powder patterns
macro > pixel³ µm³ 3D volume projection showing crystal morphology
auto All sizes Automatic method selection based on grain volume

📊 Physical Corrections

  • Lorentz factor: Proper intensity corrections for diffraction geometry
  • Polarization factor: X-ray polarization effects
  • Structure factors: Crystallographic intensity modulation
  • New scattering_factors.py module with Scherrer formula implementation

⚡ Performance Improvements

  • Batch rendering of 1M+ grains in seconds on GPU
  • Intelligent memory management with automatic batching
  • Removed NumPy dependency from hot paths

Breaking Changes

  • ScatteringUnit class removed (functionality merged into Polycrystal and Detector)
  • Rendering method names changed: centroid → micro, profiles → nano, volumes → macro
  • Minimum Python version: 3.10
  • PyTorch 2.5+ required

Testing

  • ✅ 82/82 tests passing
  • New end-to-end tests for powder diffraction with pyFAI peak finding
  • New crystallite size broadening validation tests
  • CUDA device configuration tests

Files Changed

  • 100 files modified (+10,139 / -5,764 lines)
  • Core modules refactored: detector.py, polycrystal.py, motion.py, mesh.py, laue.py
  • New modules: cuda.py, scattering_factors.py
  • Documentation rebuilt with Sphinx

Dependencies Updated

torch>=2.5.0
numpy>=1.24.0
scipy>=1.11.0
pandas>=2.0.0

Commits (20)

  1. Vectorization merge and initial refactoring
  2. PyTorch backend implementation through _diffract
  3. Torch/NumPy equivalence achieved for full simulation pipeline
  4. Multi-frame rendering support
  5. Powder diffraction mode with PSF convolution
  6. Lorentz, polarization, and structure factor corrections
  7. Memory-efficient batch processing
  8. Removed NumPy from computations
  9. Volume projection for macro grains
  10. Gaussian interpolation for micro grains
  11. Airy disk patterns for nano grains
  12. Auto mode with grain-size-based method selection
  13. Vectorized RigidBodyMotion for batch operations
  14. End-to-end powder diffraction tests
  15. CUDA device configuration module
  16. Scherrer formula implementation
  17. Documentation and docstring updates
  18. CI/CD workflow updates for Python 3.13
  19. Repository cleanup
  20. Sphinx documentation rebuild

Ready for review and merge into main.