Distribution-Shift-Aware Conformal Prediction for Meteorite and Asteroid Reflectance Spectroscopy: Finite-Sample Class Guarantees from the Laboratory to the Telescope

Abstract Machine-learning classifiers trained on laboratory reflectance spectra of meteorites are increasingly used to interpret asteroid observations, but they are routinely deployed outside the conditions under which they were validated: telescope targets differ from laboratory powders in grain size, viewing geometry, space weathering, class abundances, and sensor range. We introduce distribution-shift-aware conformal prediction to planetary reflectance spectroscopy, converting any spectral classifier into a set-valued predictor with finite-sample coverage guarantees, and we measure exactly how those guarantees behave when the deployment assumptions fail. From 16,291 RELAB Planetary Data System products, we curate 2,949 meteorite spectra into a leakage-controlled, provenance-tracked data set spanning 9 classes and 756 physical specimens. Same-specimen controlled experiments show that empirical coverage degrades with physical shift severity: grain-size mismatch removes up to 22 points of coverage, pulsed-laser space weathering removes 15 points on identical rocks, viewing geometry up to 40–60° phase leaves coverage intact, and a linear domain-classifier AUC tracks severity while proving necessary but not sufficient for failure. Class-conditional calibration is invariant to arbitrary class-prior changes and rescues the scientifically critical CI class, which marginal calibration silently abandons. Likelihood-ratio-weighted conformal prediction restores nominal coverage under a pure covariate shift but misfires when the shift carries a label component, yielding a practical combined prescription. Transferred to 694 MITHNEOS near-Earth-asteroid spectra, the calibrated sets remain informative and recover the Itokawa and Vesta-family ground truths, whereas returned-sample tests on Ryugu and Bennu expose exactly where representation, not calibration, becomes the binding constraint.