Simulate and reconstruct tomography data

Use synthetic data to check your installation and understand the array and geometry conventions before processing a scan. Follow installation first; all commands below run from the checkout root.

Generate a parallel scanCopy Anchor Link

uv run --no-sync tomojax simulate -o synthetic.nxs --size 32 --n-views 60
uv run --no-sync tomojax recon synthetic.nxs -o recon.nxs
uv run --no-sync tomojax inspect recon.nxs --preview previews

--size 32 makes a 32³ volume and a 32×32 detector. The default phantom is Shepp–Logan. Simulation writes geometry and projection metadata into the dataset, and inspect writes the central projection and the volume's central slices to previews/. Use tomojax simulate --help for other phantoms, Poisson noise (--poisson-scale) and the random seed (--seed). Detector artifacts such as dead and hot pixels, zingers and stripes, and grids or detectors of other shapes (grid, detector), are keys of a --config TOML file; tomojax simulate --config-keys lists them.

Try tilted geometryCopy Anchor Link

uv run --no-sync tomojax simulate -o tilted.nxs --geometry lamino --tilt-deg 30 \
  --size 32 --n-views 60
uv run --no-sync tomojax recon tilted.nxs -o tilted-recon.nxs \
  --method fista --iterations 50 --tv-weight 0.005 --nonnegative --roi off
uv run --no-sync tomojax inspect tilted-recon.nxs --preview tilted-previews

The tilt is measured away from the nominal tomography rotation axis. Laminography has incomplete angular information; changing the solver does not restore that missing information. Iteration count and TV weight above are starting settings, not validated choices for every object or noise level.

Use the Python APICopy Anchor Link

uv run --no-sync python examples/simulate_and_reconstruct.py

The complete example constructs a Grid, Detector and ParallelGeometry, simulates a scan with tj.project, and reconstructs it with tj.reconstruct(scan, "cgls"):

import numpy as np
import tomojax as tj
from tomojax.datasets import shepp_logan_3d

grid = tj.Grid(32, 32, 32, 1.0, 1.0, 1.0)
geometry = tj.ParallelGeometry(grid, tj.Detector(32, 32, 1.0, 1.0), np.linspace(0, 180, 60, endpoint=False))
scan = tj.Scan(tj.project(geometry, shepp_logan_3d(32, 32, 32)), geometry)
volume = tj.reconstruct(scan, "cgls", iterations=24).volume

The method and keywords of tj.reconstruct match the tomojax recon options: "cgls", iterations=, tv_weight= and nonnegative= are --method cgls, --iterations, --tv-weight and --nonnegative.

Python volumes use (x, y, z) and projections (view, v, u). All spacings use one consistent physical length unit; projected values integrate attenuation over that length. On-disk volumes may use a different labelled axis order.

This example deliberately uses matched simulation and reconstruction models. For independent analytic fixtures and failure-inclusive comparisons, see the measurement guide. For the README figure and its settings, see example reproduction.