midas-defect

Differentiable diffuse-scattering defect metrology layered on top of standard far-field HEDM.

The standard MIDAS chain (midas_index → midas_fit_grain → midas_process_grains, with midas_transforms for the detector model and midas_stress for orientations) operates on detected Bragg peaks and returns grains. It discards everything between and around the peaks. For a deformed material that diffuse field is not noise — it is the defect signal:

• Asterism — extended, anisotropic intensity hugging each Bragg core, encoding the dislocation strain field and orientation gradient. Quantifying it gives a per-grain dislocation density (Williamson–Hall on radial breadth).
• Rods in q-space — 1-D streaks threading several Bragg shells along low-index directions, the reciprocal-space signature of planar defects (stacking faults, twin walls). For FCC, faults on {111} ⇒ ⟨111⟩ rods.
• The full intensity budget — classifying every above-threshold voxel against the predicted reciprocal lattice of the indexed grains lets the scattered intensity be decomposed and closed at 100 % (Bragg / near-Bragg asterism / inter-Bragg diffuse / low-q halo). Closure is what licenses quantitative statements ("31 % dislocation asterism, 4 % fault rods") instead of qualitative ones.
• Selection-rule / forbidden-reflection tests and 3D-ΔPDF — phase and defect-symmetry checks (no second phase, no anti-phase boundaries).

This package is the productized, tested form of that whole-field, post-indexing diffuse layer.

Phase-agnostic by design. Every analysis module is driven by a midas_hkls.Crystal + a Geometry; nothing is hard-wired to FCC. The demk default is lattice.fcc_cu_crystal(), but lattice.cual2_crystal() (tetragonal θ-Al₂Cu) is retained as a first-class phase — a future CuAl₂ sample that genuinely shows superstructure rods is precisely the case the rod / forward-sim / ΔPDF machinery was built for, and re-points by swapping the crystal cell.

Shippable notebooks. midas_defect/examples/ holds end-to-end notebooks parameterized by (Crystal, Geometry, data). The flagship reproduces the demk FCC defect inventory (index → budget → ρ → rods → forbidden) and is runnable by external collaborators; the same notebook re-points to a CuAl₂ dataset by editing one cell.

See implementation_plan.md for the analysis pipelines, the MIDAS reuse map, and the test/CI architecture.

Status

Re-scoped 2026-05-21 after the demk FCC re-analysis. The package's original v0.1 framing — "heavily-deformed single crystal, treat the pattern as a continuous field because the standard pipeline cannot index it" — is retired: on the driving dataset the standard pipeline indexes cleanly (~250 grains/layer), the material is an ordinary FCC polycrystal/sub-grain mosaic, and the diffuse field is dominated by dislocation asterism (~31 %) over rods (~4 %). The corrected scope is diffuse defect metrology on top of, not instead of, standard FF-HEDM.

Pre-alpha. v0.1 target: shared infrastructure (geometry, lattice, bragg_diffuse) + the three capabilities that close the budget (intensity_budget, williamson_hall, defect_tests), each validated to reproduce the published demk numbers. Rod/asterism/sub-grain/ΔPDF/forward-sim modules exist and pass synthetic tests; they promote once anchored to real-data regressions.

Module status

The validated core (v0.1) is anchored to the published demk full-res numbers; the rest is synthetic-tested and promotes once anchored to real data.

defect_tests.rod_family_enrichment is a screening metric only — confounded by reciprocal-lattice geometry; use fault_rod_alignment (explicit per-grain) for the authoritative ⟨111⟩ fault-rod test.

Driving dataset

Sep-2025 1-ID-E beamtime — 10 Y-layers × 1440 ω-frames @ 0.25°/frame on Pilatus3 CdTe 2M, 71.676 keV (λ = 0.172979 Å). Phase: FCC Cu(-rich solid solution), a = 3.6356 Å (space group 225). θ-Al₂Cu (CuAl₂, I4/mcm, a ≈ 6.066 Å) was the a-priori candidate given the Cu–Al provenance and was tested and eliminated from the powder line-out (the 1/d² ring sequence is the exact FCC fingerprint; CuAl₂'s strongest (110) line at d = 4.29 Å is absent). The validated gold-calibrant detector geometry (Lsd 652.7 mm, tilts + distortion, correct ω sign) is the default in geometry.demk_default_geometry().

Published analysis (the worked validation case for this package): ~/Desktop/analysis/demk/fcc_reanalysis/ (FINDINGS.md + figures + scripts).

Engineering contract

All modules satisfy ALL four:

1. Differentiable. Every physics / scoring / fitting routine is torch; numpy / scipy live only in _discrete helpers off the gradient path.
2. Device portable. CPU / CUDA / MPS via midas_transforms.device.resolve_device.
3. Thoroughly tested. Per module: synthetic + autograd + device + real-data regression (reproduces a published demk number) + benchmark.
4. Reuses upstream MIDAS. No re-ported orientation, lattice, transform, or device-resolution code. See the reuse map in implementation_plan.md.