midas-uq

Cross-validation based uncertainty quantification for High-Energy Diffraction Microscopy (HEDM) grain refinement. Three diagnostics, three modalities (far-field, point-focused, near-field), all built on the end-to-end differentiable forward model of midas-diffract.

Companion paper: H. Sharma, J.-S. Park, P. Kenesei, N. Andrejevic & M. Cherukara, Cross-Validation Based Uncertainty Quantification for HEDM Grain Refinement, IUCrJ (in preparation, 2026).

What it does

For each grain in a polycrystal HEDM refinement, midas-uq answers:

• How reproducible is this grain's refined state under random resampling of its measured spots (or omega frames)? — half_half
• Which individual measurements are driving the fit, and which (if any) appear corrupted? — jackknife
• What is the local Gaussian-posterior covariance from the inverse Hessian? — laplace_covariance
• Is refinement overfitting in the low-spots-per-DOF regime? — rfree_gap

No ground truth required. Half-half disagreement and jackknife influence are post-hoc and label-free. The differentiable forward model from paper I makes population-scale resampling (hundreds of grains × tens of splits) a single-CPU minute, not a day.

Installation

pip install midas-uq         # adds midas-diffract automatically

Optional CPU multi-processing for population studies:

pip install "midas-uq[mp]"

Notebooks

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

Quick start (Python)

import torch
import midas_uq as muq
import midas_diffract as md

# Build the forward model the usual way (see midas-diffract docs)
geom  = md.HEDMGeometry(...)
model = md.HEDMForwardModel(hkls=..., thetas=..., geometry=geom)

# Grain seed (e.g., from a Grains.csv row)
init = muq.GrainState(
    euler_rad=torch.tensor([phi1, Phi, phi2], dtype=torch.float64),
    latc=torch.tensor([a, b, c, alpha, beta, gamma], dtype=torch.float64),
    pos=torch.tensor([x, y, z], dtype=torch.float64),
)

# Observed spots: (N, 3) of (2theta, eta, omega) in radians
obs = ...

# 1. Half-half UQ
uq = muq.half_half(model, init, obs, mode="ff", n_splits=5)
print(uq.misori_median_deg, uq.lattice_median_A)

# 2. Per-spot jackknife (only on a flagged grain)
jk = muq.jackknife(model, init, obs, mode="ff")
print(jk.top_k(10))  # 10 most influential spots

# 3. Laplace baseline for comparison
sigma_vec = torch.tensor([sigma_2theta, sigma_eta, sigma_omega],
                         dtype=torch.float64)
lp = muq.laplace_covariance(model, init, obs, sigma_vec, refine_first=True)
print(lp.misori_p95_deg, lp.condition_number)

For NF-HEDM, observations is a (F, H, W) detector volume and the API is the same with mode="nf"; see examples/nf_frame_split.py.

Quick start (CLI)

For a standard MIDAS dataset (Grains.csv + SpotMatrix.csv + paramstest + hkls.csv):

# Population half-half UQ (writes one row per grain)
midas-uq half-half \
    --params  /path/to/paramstest.txt \
    --hkls    /path/to/hkls.csv \
    --grains  /path/to/Grains.csv \
    --spot-matrix /path/to/SpotMatrix.csv \
    --n-splits 5 \
    --out uq_half_half.csv

# Drill into a single grain
midas-uq jackknife --grain-id 5823 ...
midas-uq laplace   --grain-id 5823 ...

API surface

Result dataclasses: HalfHalfResult, JackknifeResult, LaplaceResult, RFreeResult.

When to use what

Half-half is the recommended population-scale diagnostic; jackknife is the drill-down on grains flagged by half-half; Laplace gives a complementary Gaussian baseline whose discrepancy from the empirical half-half spread itself diagnoses non-Gaussian posterior structure.

Reproducing the companion paper

The Ti-7Al population study, Park22 in-situ tensile sweep, synthetic sweeps, and figure scripts live in dev/paper/ of the repository.

License

BSD-3-Clause.