ReflectoSim

ReflectoSim generates realistic synthetic X-ray reflectivity (XRR) curves for training and testing inverse models. You describe a distribution, a prior over layer stacks and instruments, and it draws curves from it with counting noise and writes them to a dataset file that records how to regenerate it.

The structure prior covers random multilayers: layer count, thickness, roughness tied to thickness (σ ≤ ratio · d), SLD with a minimum contrast between neighbours, absorption, and a native oxide whose position you must state (on the substrate, on top, or none). The instrument prior covers resolution, 2θ offset, I0, background, Cu Kα₂, footprint and counting statistics. Counts are Poisson, with optional dark counts and read noise.

The curves come from the forward model of ReflectoFit, the fitting engine; ReflectoSim adds the data distribution and has no physics of its own. Before noise, every curve equals reflectofit.forward for its row, a curve does not depend on how the work is chunked, and the same seed gives the same file.

Install

Not on PyPI yet. From a clone, with uv (it fetches ReflectoFit from GitHub):

git clone https://github.com/NaNIsaac/ReflectoSim
cd ReflectoSim
uv sync

First run

Simulate 1000 curves of one to three films on SiO₂ on Si from the example prior and summarise the file:

uv run reflectosim sim --prior examples/lab_films.toml --n 1000 --seed 20261024,1 -o train.npz
uv run reflectosim describe train.npz

The dataset layout, the prior file format and the Python API are in the README.

Accuracy and cost

By default the resolution integral is converged to about 1e-5 (ReflectoFit's per-point rule), which is the slow part for wide resolution and weakly absorbing films. InstrumentPrior.nodes fixes a smaller node count for speed; per NaNIsaac DECISIONS #24 that is for generators who state an error budget against the counting noise in advance and record it in the dataset metadata.