Published applications

These checks run published scientific workflows with treams-rs and compare the results with the authors' numerical data or fresh runs of their examples. They complement tests of individual numerical functions; agreement alone does not prove accuracy. Every comparison names its source, sampling grid, truncation order and tolerance. Runtimes in these results are diagnostics, not controlled benchmarks.

Lines use the full original notebook grids; circles show the separately sampled author reference tables. Narrow cylinder resonances explain the visual difference between line interpolation and the sparse points. Both grids are evaluated and checked separately. The cylinder notebook and regression table also use the slightly different dispersion constants recorded below.

Electron-beam spectroscopy

Stamatopoulou and Rockstuhl, A T-matrix scattering formalism for electron-beam spectroscopy, arXiv:2602.12743v1 (2026), Figure 2(a,b).

The public author repository provides sphere/cylinder notebooks and stored regression spectra. treams-rs evaluates the electron source, cylindrical translation, cylindrical-to-spherical conversion, Mie response, and cathodoluminescence (CL) / electron energy loss (EELS) pairings. The observable formulas follow paper equations 6, 21, 25, and 26. The same explicit calculations also run with treams==0.4.5.

Both particles have radius 50 nm, electron impact parameter 60 nm, and reduced speed 0.7. The sphere has permittivity 16+0.5j and order 4; the infinite cylinder uses the specified silver-like Drude model and order 12. Both full 200-point notebook grids are reproduced. Separately, 50 points per object are checked against the authors' stored test values (200 CL/EELS values). These test grids differ from the notebooks. The cylinder notebook's rounded hbar=6.582e-16 eV s is preserved in its dispersion model; the stored-test calculation uses SciPy constants.

Running the authors' notebook calculations exposed a cylinder-source error: the notebook omits the conversion from the helicity T-matrix to the parity electron-source basis. Its own regression test includes that conversion. The check applies that one-line repair explicitly; the unmodified notebook differs by as much as 50% in EELS. The unmodified sphere notebook and repaired cylinder notebook match treams-rs within 8e-17 absolute. Both original and repaired results are preserved in ebeam-notebook-execution.json. The plotted cylinder curve uses the corrected notebook calculation.

The independent stored-data tolerance is atol=1e-8, rtol=0, matching the precision of the rounded source tables. Maximum error is 4.84e-9 for the sphere and 2.66e-12 for the cylinder. Fresh treams calculations agree much more closely. CL and EELS are nonnegative, CL does not exceed EELS, and rotational invariance of the sphere is checked with Hypothesis. Cylinder CL vanishes because every axial channel is evanescent. Increasing the multipole cutoff by two at five frequencies changes sphere EELS by up to 0.24% and cylinder EELS by up to 0.62%; reproducing the author's truncation is not a claim of arbitrary-precision convergence.

Sources: sphere notebook, cylinder notebook, sphere data, cylinder data. Only numerical reference values are stored locally, with source hashes and attribution; the addon's implementation is not included.

CPC treams paper companion spectra

Beutel, Fernandez-Corbaton, and Rockstuhl, treams – a T-matrix-based scattering code for nanophotonics, Computer Physics Communications 297, 109076 (2024).

Table 3 refers readers to accompanying executable examples. treams-rs reproduces the 200-point dielectric sphere efficiency spectrum, 50-point absorbing chiral slab transmission/reflection spectra, and periodic chiral spheres above a dielectric slab. Parameters and grids come from the treams examples at the reference version: sphere, slab, periodic array. These results reproduce the companion examples; they were not read from the article's plots.

All checked samples agree with fresh treams values within rtol=2e-9, atol=2e-11. Passive sphere/slab inequalities hold; the lossless periodic stack conserves T+R=1 within 1e-10. The original array order 3 differs from order 5 by up to 1.1e-4 in transmittance at the five convergence samples.

The array's final original sample, wavelength 350 nm, is an exact diffraction threshold and is recorded as excluded; the 99 interior samples are checked. treams assigns finite surrogates to both the radiation pole and the lattice pole. That finite output is not an independent exact-threshold reference. Four extra evaluations at 350*(1±1e-4) and 350*(1±1e-6) nm verify agreement away from the pole, power conservation, and convergence of the two sides. The original endpoint remains excluded, and its value and derivative are not checked.

The article's quasi-BIC Figure 5 is not reproduced here. Appendix F supplies the algorithm, but ellipsoid.h5 is absent from the retrieved arXiv ancillary archive. The recorded retrieval of the publisher supplement returned HTTP 403. Replacing the unavailable FEM matrix by another ellipsoid model would be a different check, so no Figure 5 claim is made.

Thermal radiation

Mazo-Vásquez et al., Studying thermal radiation with T matrices, Physical Review B 112, 054307 (2025), Figure 2(a,c). The original notebook and data describe four 250 nm SiC spheres separated by 520 nm, with a temperature of 500 K. Their material data and all 300 original frequencies are retained. At multipole order 10, the local interaction system has 960 channels and is expanded into a 240-channel global basis. treams-rs reproduces the absorption spectrum and three complete 200-angle thermal-emission cuts against the authors' original numeric tables.

All 300 absorption values agree within 0.577%; all 600 angular values agree within 2.18e-7 relative. The tolerance for the absorption reproduction is 1%, consistent with the paper's stated 99% multipole-convergence criterion. Increasing the cutoff through order 12, and through order 16 at the lowest frequency, changes order-10 absorption by at most 0.057% over 22 checks. Positive absorption and azimuth/helicity symmetry are verified; the maximum symmetry discrepancy is 1.56e-14. A separate 52-point comparison with treams is kept.

Small particles at high orders make the dense interaction system badly scaled. Scaling the interaction system before solving removes this sensitivity: at the lowest frequency the stable order-10 value is 1.7544335861e-4 µm², while the author table gives 1.7464535052e-4 µm². This disagreement remains in the results. The original angular filenames use another figure numbering; the source-data file records their mapping, units and the notebook's file-saving convention.

Results: thermal-result.json, source data and hashes, treams comparison.

uv run --no-sync python scripts/papers_thermal.py --samples 300 --convergence \
  --upstream-reference benchmarks/papers/thermal-upstream-comparison.json \
  --output benchmarks/results/local/thermal-result.json

This command recomputes every native sample and reuses the treams values of the committed comparison, which has the same parameters. Omit --upstream-reference to recompute treams for all 300 frequencies too. Source material is fixed to the cited commit; the command uses local reference data by default. The checks cover only the selected figures.

Reproduce and inspect

From a built checkout:

uv run --no-sync python scripts/qualify_papers.py \
  --output benchmarks/results/local/qualification.json
uv run --no-sync pytest -q tests/scripts/test_paper_qualification.py
uv run --no-sync --with matplotlib python scripts/qualify_papers.py --plot \
  --output benchmarks/results/local/qualification.json

The script checks full spectra, physical inequalities and source-data tolerances, and writes its result to the given --output file; with --plot it also writes PNG and SVG figures next to it. The committed result is qualification.json. The JSON includes all grids and values, package versions, and the installed native binary and script SHA-256 hashes. Source URLs and hashes are in source-provenance.json. The compact author reference tables are exercised by ordinary tests, so the reference checks do not require network access. The quick tests in tests/scripts/test_paper_qualification.py include resonance samples, two original thermal absorption samples and a Hypothesis test of rotation invariance and passivity.