ReflectoExp

Experiment record. The maintained library, with on-the-fly simulation and a differentiable Abeles engine, is ReflectoLearn.

Experiment scripts for training neural networks that predict thin-film thickness, roughness and SLD from X-ray reflectivity (XRR) curves, using GenX (and earlier refnx) simulations as training data.

Each folder under runs/ is one self-contained experiment (config, simulation, dataset, model, training, evaluation). Shared simulation and physics helpers live in the reflecto_exp package under src/.

Install

Uses uv and Python 3.13. torch is pulled from the PyTorch CUDA 12.8 index.

git clone https://github.com/NaNIsaac/ReflectoExp.git
cd ReflectoExp
uv sync
uv pip install -e .     # installs the reflecto_exp package

The refnx-based experiments (exp02, exp03) also need uv pip install refnx, which is not a declared dependency.

Usage

Experiment scripts import their siblings as top-level modules (import config, import simulate), so they are run from inside their folder. The root main.py does that:

uv run python main.py 7      # runs runs/exp07_fourier_hires/main.py with the .venv Python

main.py of an experiment generates the HDF5 training set if it is missing, then trains and evaluates. Before running, set base_dir (and exp_name) in that experiment's config.py; the committed values are local Windows paths.

Experiments

Folder	What it tries
exp01_quan_class	multilayer parameters as quantized classes (imports modules that no longer exist in reflecto_exp; kept for reference)
exp02_one_layer	single-layer regression on refnx simulations
exp03_physics_fused	CNN prediction blended with a physics-based thickness estimate, experimental
exp04_one_genx	single-layer regression on GenX simulations
exp05_1layer_mask	single layer, valid-q mask as a second input channel, instrument augmentation; GenX fit started from the CNN guess
exp06_1lay_6out	film + SiO2 native oxide (6 outputs)
exp07_fourier_hires	6 outputs, 2000-point q grid, Fourier features of q; ablation.py and metrics_report.py for component ablation and extended metrics

Structure

src/reflecto_exp/
  simulate/simul_genx.py      GenX-based XRR simulator (XRRSimulator, ParamSet)
  simulate/simul_refnx.py     refnx-based simulator
  simulate/noise.py           Poisson and background noise
  simulate/torch_engine/      PyTorch Abeles / kinematical reflectivity, smearing
  physics_utils.py            unit conversions (tth2q, ...)
runs/expNN_*/                 experiments
notebooks/                    function tests

Citation

See CITATION.cff.

License

MIT, see LICENSE.