midas-invert 0.1.0

Domain-agnostic differentiable-inversion primitives for MIDAS: gradient fitting, scale-invariant losses, Laplace uncertainty, Fisher-information experiment design, mixture deconvolution, and amortised-inference surrogates.

midas-invert

Domain-agnostic differentiable-inversion primitives for MIDAS, shared across HEDM, Laue microdiffraction, pf-/grain-ODF, and 2D/ultrafast diffraction. None of it knows about diffraction — you supply the forward model and the loss closure.

• optimize — fit (Adam / L-BFGS), relative_l2_loss (scale-robust), cosine_loss (scale-invariant shape loss; avoids the argmax kink of peak-normalisation — right for rocking curves, fringe profiles, spectra).
• uq — laplace_uncertainty (Hessian-at-optimum covariance & std-devs).
• design — fisher_information, rank_measurements, next_best_measurement (which delay / reflection / energy best constrains a target — experiment design).
• mixture — mixture_deconvolution (recover non-negative softmax weights over a component grid: thickness/grain-size distributions, ODF on a grid, a spectrum over an energy grid).
• surrogate — ParameterMLP, train_surrogate (amortised inference; the differentiable forward is the data generator).

All torch-differentiable, CPU / CUDA / MPS. Extracted from midas_2d so HEDM, Laue, and 2D draw on one implementation (see ../HEDM_LAUE_TRANSFER_PLAN.md).

pip install -e . --no-deps
KMP_DUPLICATE_LIB_OK=TRUE pytest