Reflecto

Single-layer X-ray reflectivity (XRR) analysis: a 1D CNN predicts film thickness, roughness and SLD from a measured curve (2θ, intensity .dat), then GenX refines a Si / SiO₂ / film model (native oxide under the film, the stack of the training data) starting from that guess.

The bundled weights (src/reflecto/assets/best.pt, stats.pt, config.json) come from experiment exp07 in ReflectoExp as of December 2025: config.json has the values of exp07's config.py at ReflectoExp b93c91f (2025-12-23: exp_name "cubic", fixed SiO₂ SLD 18.8, roughness 0-15 Å), not those of its current config.

Reflecto is one of the XRR research branches (CNN guess, then GenX differential evolution). ReflectoLearn carries an adapted copy of the GenX fit as reflectolearn.genx_fit (ReflectoLearn 6e160db). No comparison of the two approaches has been run, and the electron-density profile and density-ratio output here have no counterpart there.

Usage

Python 3.13 or newer with uv. pyproject.toml has no build system, so the package is not installed and src goes on the path. From the repository root (PowerShell):

uv sync
$env:PYTHONPATH = "src"
uv run main.py path\to\curve.dat

main.py analyses one measured curve (two columns, 2θ and intensity) of an In₂O₃ film: CNN prediction, GenX fit, the density ratio against bulk and the figure of merit, then plots of the reflectivity curve and the electron-density profile. For another material, change the In₂O₃ constants at the top of main(). No measured data ships with the repository.

import reflecto
from reflecto.utils.math_utils import i0_normalize

df = reflecto.read_dat("sample.dat")  # two columns: tth, intensity (whitespace, # comments)
q = reflecto.tth2q(df["tth"].to_numpy())
R = i0_normalize(df["intensity"].to_numpy())
result = reflecto.analyze(
    q, R, fit=True
)  # result["prediction"], result["fit_params"], result["fom"]
reflecto.plot_analysis_result(q, R, result)

Structure

src/reflecto/
  core.py            analyze(): CNN inference and GenX fit
  engines/           CNN model, inference (inference.py), GenX fit (fitting.py, script_builder.py)
  simulate/          GenX reflectivity simulation, noise, SLD profiles
  utils/             .dat reading, 2θ to q, SLD and density conversion, GenX layer constants
  assets/            trained weights (best.pt, about 64 MB)

Dependencies: numpy, torch, genx3, bumps, pandas, matplotlib, tqdm (pyproject.toml, uv.lock).