Uncertainty-aware machine learning for core-loss background subtraction in EELS.

Quantitative electron energy-loss spectroscopy (EELS) critically depends on accurate background subtraction, yet standard approaches rely on deterministic parametric fits and provide no statistically rigorous uncertainty estimates. This limitation is particularly severe in the core-loss regime, where the background must be extrapolated into the ionisation-edge window, hindering reliable and automated analyses. Here, we introduce a physics-informed, uncertainty-aware machine learning framework that reformulates core-loss background subtraction as a statistically controlled inference problem with explicit uncertainty propagation. The method combines Monte Carlo replica generation based on empirically determined covariance matrices with neural network ensembles trained on the pre-edge region, and enforces physically consistent extrapolation via an effective-exponent constraint at the edge onset. This approach yields pixel- and energy-resolved background probability distributions, enabling direct propagation of uncertainty to derived spectroscopic observables. Closure tests on synthetic spectral images demonstrate faithful background reconstruction and well-calibrated uncertainty estimates, including in extrapolation regions. Applied to core-loss EELS spectral images of twisted CrSBr, the framework reveals nanoscale, stacking-correlated modulations of Cr L2,3 white-line intensities that significantly exceed propagated uncertainties. More broadly, this work provides an automation-ready pathway for quantitative, uncertainty-aware core-loss EELS analysis across diverse materials systems.