Deep learning AI reconstructs 3D chemical maps from limited-angle STEM-EDX

Researchers from CEA-Leti and Université Paris-Saclay developed an unsupervised deep learning framework called DIPm-TV (multi-channel Deep Image Prior with total variation regularization) for limited-angle STEM-EDX tomography. EDX tomography in scanning transmission electron microscopy (STEM) provides 3D elemental mapping at the nanoscale, but is severely limited by restricted tilt ranges and low-dose conditions to avoid beam damage. The missing-wedge problem (here ~100° of missing data) introduces elongation and anisotropic resolution artifacts, often making conventional reconstruction methods like SIRT or compressed sensing inadequate.

The team extended the Deep Image Prior approach to a multi-channel formulation that jointly reconstructs multiple elemental maps by exploiting spatial correlations. They validated the method on a synthetic 3-channel phantom and then applied it to real Ge-Sb-Te (GST) phase-change memory devices. Samples were imaged from -40° to +40° with 5° steps and a dose of 2.0×10^5 e⁻/Ų. The method achieves near-isotropic spatial resolution and reveals compositional heterogeneities associated with device operation (virgin vs. SET crystalline states). This approach enables 3D chemical characterization in experimentally accessible sample geometries where conventional methods fail, offering a practical path for nanoscale materials science.