midas-propagate 0.1.1

End-to-end uncertainty propagation and joint re-refinement for HEDM (calibration -> indexing -> per-grain refinement -> stress)

midas-propagate

Calibration-aware uncertainty propagation and joint re-refinement for end-to-end HEDM grain analysis.

Status: scaffold. See dev/paper/SKETCH.md for the paper roadmap. No working code yet.

Notebooks

Worked-example Jupyter notebooks live in notebooks/. They are not shipped with pip install — get them by cloning the MIDAS repository.

What this package will do

1. Compose the existing differentiable losses across the four pipeline stages — calibration (midas-calibrate-v2), indexing (midas-index), per-grain refinement (midas-fit-grain), and elastic inversion (midas-stress) — into a single joint NLL.

2. Compute a joint MAP estimate over any subset of:

   • detector calibration (Lsd, tilts, beam center, distortion)
   • per-grain orientation, lattice strain, position
   • global nuisance parameters (wavelength, beam profile)
   • optionally per-grain elastic stress

3. Assemble the block-structured Hessian at MAP and return calibration-aware per-grain covariance via Schur-complement marginalization — no full-matrix inverse needed.

4. Propagate per-grain covariance through the elastic inversion to per-grain stress error bars via the delta method.

Why

Production HEDM tools (HEXRD, MIDAS, ImageD11, FABLE) report per-grain σ at the converged grain state with detector calibration held fixed. Downstream Bayesian crystal-plasticity work (Greeley 2026, Iyer 2025) explicitly assumes an HEDM σ no current tool can derive. This package closes that loop.

Companion packages

• midas-diffract — differentiable HEDM forward model
• midas-uq — single-grain holdout / jackknife / Laplace UQ
• midas-joint-ff-calibrate — multi-grain + multi-detector joint refinement
• midas-calibrate-v2 — instrument calibration
• midas-stress — single-crystal elastic inversion