Repository navigation
EDX fit accuracy on real Al alloys #8
Description
Mg Kα about 300 counts low, Si 8.7 % low, σ 5 to 12 % small; Mg missed in 37 of 38 Al-Mg-Si files. Diagnose on the Al matrix of a real Al-Mg-Si map (pixels whose Mg and Si are below L_D) as the in-situ standard. Known issue: offset and gain trade off on lines that span little of the axis (AxisGateTests.testE3RefinementRecoversPlantedOffsetAndGain).
From docs/open-items.md item 3, 27 Oct 2026.
Activity
Paullo9 commentedon Oct 31, 2026
Diagnosis done locally on the owner's data; the data and every number measured on it stay private. No kit change in this round: no single kit defect with a clear fix was found.
What the experiment separated, in general terms:
eds-absorption-truth.json(no detector Si edge) the same split triples the Si sigma and moves Si by -4 % to +19 % depending on the order, so it is not shipped as a default.FitPolicy.continuumEdges/continuumOrdersalready let a consumer choose it.Next: a physical detector Si-edge term in the continuum (no free polynomial freedom under Si K-alpha), tested on a planted spectrum that has a dead-layer step, then measured on a Si-free standard. All results stay
validation: "none".test: golden baseline re-recorded for continuumDeadLayerMicrons, no n…
Paullo9 commentedon Oct 31, 2026
Checked on the public AlMgSi demo set (
demo-edx, generatorsim_edx.pyv2.0, seed 42, ladder seed 52), which carries planted truth on synthetic data: recipe compositions, the true k, four-segment absorption, a 1.5 % Al K-alpha tail (0.12 keV), Si internal fluorescence, Si escape, sum peaks (parents lose the photons) and stray Cu K. No detector dead layer is planted.Pipeline:
PooledQuantifier.runwith the true k typed, the truth geometry and thickness, absorption on, the true axis locked. Each pool was fitted twice: the Poisson data, and the noise-free expected spectrum x100 (the fit bias without noise). Si planted counts include the Si internal-fluorescence peak (the kit convention).Default fit (exSpy-equivalent: one Gaussian per line, no sum peaks), net against planted, noise-free:
eds-absorption-truth.json(recipe A, 1e4 px)At% on the data: matrix map Mg 0.44 / Si 0.45 against 0.60 / 0.60 truth; L5 Si 0.36 against 0.60.
The Si deficit follows the unmodelled sum peaks, not a dead layer. The Al+Al sum peak (2.97 keV) and the other sum peaks pull the continuum segment above the Al K edge; under least squares that lowers Si K-alpha. The cheap refuting experiment was to switch on
SumEscapeModel(sumPeaks: true)and see whether Si stays biased. It does not: noise-free Si goes from -18.2 % to -4.0 % on the fixture, -16.6 % to -3.9 % on the matrix map, -14.9 % to -3.6 % on recipe A, -57 % to -8 % on L5 and -8.4 % to -1.2 % on L6. Reduced chi2 on the x100 spectra falls from 120-3000 to 0.04-8.The earlier diagnosis on real data placed the Si deficit in the continuum between the Al K edge and 2.2 keV. On synthetic data the deficit follows the sum peaks instead, so the dead layer and a Si-free standard are no longer the next step here.
Dead-layer term (#50). A typed 0 um reproduces the default fit exactly. Fitted on this detector, which has no dead layer, it reads 0.17-0.43 um and moves Si further down in every Al pool: noise-free matrix map from -16.6 % to -23.8 %, L5 from -57 % to -75 %. On the pure beta'' pool it is neutral. The fitted thickness absorbs the sum-peak misfit, so it should be fitted only together with the sum-peak model, or typed.
Absorption correction: right. On L7 the corrected Mg/Si pull is -0.4 and the uncorrected one 5.0.
Next step: the pooled path models sum peaks when the data call for it, keeps the exSpy-equivalent fit as a switch, and pins these demo numbers as regression tests (PR to follow). Results stay
validation: "none": the truth is synthetic.Paullo9 commentedon Oct 31, 2026
The continuum at the Al K edge on a real Si-free spectrum: diagnosis and #64
Data: Snelson, Collins and MacArthur (2026), Elemental tomography dataset of Mg-Al-Ca alloy, University of Leeds, https://doi.org/10.5518/1826, licence CC BY 4.0. File
Al_Wedge_Standard/Al_EDS-HAADF.emd: Titan G2 at 200 kV, Super-X. Pool of 214 793 px, 3e5 Al K-alpha counts, the refined axis from #15.Cause. There are two parts.
What would refute part 1: broadening the step leaves the misfit unchanged. It does not; it halves it.
exSpy 0.3.2 has no edge in its background: it fits a polynomial (order 6) across the whole range. The step is the kit's own design, and no parity path is involved.
The detector dead layer (#50) is not the lever here. The data show no step at 1.839 keV, and the 10 nm Al contact layer of the generic SDD model gives about a 1 % step at the Al edge.
Synthetic check, generated in the test. An Al film whose edge is convolved with the detector, 3e5 Al K-alpha, weak Si K-alpha 2000 counts, noise-free:
The AlMgSi demo (sim_edx.py v2.0, noise-free pools ×100, true axis; planted truth on synthetic data) plants an unbroadened step:
#64, stacked on #59, adds
FitPolicy.continuumEdgeStep = .broadenedas an option..sharpstays the default and is the earlier form, value for value. Making it the default first needs the kit's registered synthetic generators moved to a convolved continuum: five registered results move, because they plant a sharp step.Next on this issue:
Ice layer on the detector (the O question from #15), measured at no extra cost on the same spectrum. Ice was put on the generic SDD efficiency, broadened by the resolution, and the continuum fitted against it:
It is a first number, not a calibration.
Results stay
validation: "none".🤖 Generated with Claude Code
feat: detector Si dead-layer step in the EDX simulator and a planted …
feat: optional detector Si dead-layer term in the EDX continuum (#50)
Paullo9 commentedon Oct 31, 2026
What is left above the Al K edge, and making the broadened step the default (#65, stacked on #64)
Real data: Snelson, Collins and MacArthur (2026), Elemental tomography dataset of Mg-Al-Ca alloy, University of Leeds, https://doi.org/10.5518/1826, CC BY 4.0. File
Al_Wedge_Standard/Al_EDS-HAADF.emd(Titan G2 200 kV, Super-X): a pool of 214 793 px on the refined axis. Synthetic truth is planted, not measured. All results stayvalidation: "none".1. The residual above the edge is the Al K-beta line
Cause: the tied Al K-beta weight. eXSpy's
xray_lines.jsonties Al Kb at 0.0132 of Ka, at 1.5596 keV. NIST EPQ gives 0.0315.What would refute it: a free line at the K-beta energy that does not take up the excess, or a χ² minimum somewhere else when the line's energy is scanned. Neither happens.
Leeds wedge, broadened step unless stated otherwise:
A free Gaussian scanned from 1.54 to 1.63 keV has its χ²/n minimum at 1.56-1.58 keV (2.34, against 2.92 at either end of the scan). That is the K-beta energy.
The other two candidates:
Synthetic Al film generated in the test (3e5 Al Ka, 2000 Si Ka, noise-free, planted Kb 0.0132). Si bias:
Demo pools with the default sum peaks, noise-free, blurred generator. Si bias:
No change ships for the residual. Each candidate fails on one side:
Next: a K-beta weight option, measured on a synthetic set that plants the EPQ weight.
2. The broadened step is the default (#65)
tools/edx-sim/detector_blur.py). Its rebuilt continuum matches forward.py to 5e-16. The demo pools are rebuilt from the sim_edx v2.0 truth arrays and checked channel by channel in 9.5-19 keV..sharpstays available; it reproduces the old default value for value."continuumEdgeStep":"broadened". A stored policy without the key reads.sharp.Leeds wedge, before and after:
Demo pools (blurred continuum, noise-free, default sum peaks), sharp → broadened:
The registered results, re-measured under the new default:
Absorption-truth Si: the pin goes from −6.7 % to −1.67 %.
Weak-line badge (A and S are the two generators), sharp → broadened:
Realistic-L: the Ti pins move by up to 13 counts.
Registered bar lowered: the proposer's 1 % Mg ×100 cell on the Kramers continuum is now a candidate in 5 of 6 draws, not 6. One draw has no residual peak at Mg.
Gate green, the parity target included. Two numerics reviews were done before the
feat!commit. CI is pending.🤖 Generated with Claude Code