Abstract
Abstract
Energy-dispersive X-ray (EDX) tomography in scanning transmission electron microscopy (STEM) enables three-dimensional elemental mapping at the nanoscale. However, its application remains challenging when sample geometry severely restricts the accessible tilt range, and the electron dose must be limited to prevent beam damage. Limited-angle acquisition produces missing-wedge artifacts, including elongation and anisotropic resolution, while low-dose measurements further degrade reconstruction quality and hinder reliable quantification. Here, we introduce an unsupervised deep-learning framework combining a Deep Image Prior with total-variation regularization (DIP-TV) for STEM-EDX tomography under extremely limited-angle conditions. We further propose a multi-channel formulation, DIPm-TV, that jointly reconstructs multiple elemental volumes by exploiting their spatial correlations. Using a synthetic three-channel phantom, we show that DIPm-TV substantially reduces artifacts arising from severe angular limitations in the presence of moderate noise, outperforming the simultaneous iterative reconstruction technique and compressed-sensing-based approaches. We then apply DIPm-TV to two Ge–Sb–Te-based memory devices: an as-fabricated virgin device and a device programmed into the SET state, both prepared as conventional cross-sectional focused-ion-beam lamellae. Under a severely limited tilt range (±40°) and a low-dose (
$$2.0\times {10}^{5}{e}^{-}{\text{\AA }}^{-2}$$
2.0
×
10
5
e
−
Å
−
2
), DIPm-TV yields reliable voxel-by-voxel 3D elemental maps from EDX signals alone, without external structural priors such as HAADF-STEM images. The reconstructed volumes exhibit near-isotropic spatial resolution and reveal compositional heterogeneities associated with device operation. This approach enables 3D chemical characterization of semiconductor devices in experimentally accessible sample geometries for which conventional reconstruction methods are strongly compromised by angular limitations.