Skip to content

OpenLSR-X Documentation

OpenLSR-X is a PyTorch Lightning package and reproducible benchmark suite for 2D-to-3D Super-Resolution of X-Ray Computed Tomography (XCT) scans of porous media.

Why OpenLSR-X?

X-ray micro-CT imaging faces a fundamental trade-off between field of view (FOV) and spatial resolution. Resolving sub-micron pore throats typically restricts sample diameters to a few millimeters—below the representative elementary volume (REV) of heterogeneous sandstones, carbonates, and catalyst pellets—while large-FOV scans blur critical pore throats and bias downstream flow simulations.

Directly training 3D convolutional or adversarial networks on gigavoxel XCT volumes requires prohibitive GPU memory and paired 3D ground-truth volumes that are costly to acquire and register. OpenLSR-X addresses both limitations:

  1. Memory-Efficient 2D Slice Training: Trains an XCT-tailored SRGAN (or non-adversarial SRResNet ablation) on 2D slices using reflection border padding, channel-wise PixelNorm, generator warmup epochs, and optional VGG-19 perceptual supervision.
  2. Isotropic 3D Orthogonal Fourier Fusion: Applies the trained 2D generator across all three orthogonal planes (yz, xz, xy) of a 3D low-resolution scan and fuses them in 3D Fourier space by weighting each view's super-resolved in-plane spatial frequencies—eliminating slice-stacking ("venetian blind") artifacts.
  3. Domain-Specific Petrophysical Validation: Quantifies accuracy on physical porous-media properties—porosity, pore-phase Dice/IoU, specific surface area, Euler connectivity, inscribed pore-radius distributions, and two-point correlation functions.

Next Steps