Accelerating 3D Semiconductor Chemistry Mapping by Automated TEM Tomography

Abstract

Three-dimensional (3D) chemical characterization at the transistor level is increasingly required for root-cause failure analysis (FA) of advanced semiconductor devices, whose yield-limiting defects are intrinsically volumetric. Multimodal electron tomography combining high-angle annular dark-field (HAADF) scanning/transmission electron microscopy (S/TEM) with spectroscopic chemical signal, such as energy-dispersive X-ray spectroscopy (EDX) or electron energy-loss spectroscopy (EELS), is the established route to nanometer-resolution 3D elemental volumes. The conventional workflow has remained impractical for routine FA: it relies on needle-shaped specimens, dedicated on-axis tomography holders, tilt ranges of ±70° or wider, dense projection sets, fiducial-based alignment, and a sequential pipeline of manual pre-processing, reconstruction, and post-processing that consumes most of a working day per dataset. We present a self-supervised reconstruction framework that recasts multimodal electron tomography as a single inverse problem: a differentiable physical forward model is coupled with a learnable volume representation, and the 3D chemical volume, tilt-series alignment, detector response, and per-element intensity scaling are recovered jointly from raw projections without external training data, prior alignment, or per-sample tuning. Unlike classical reconstruction algorithms, the framework encodes the imaging physics of both modalities and exploits their cross-modal coupling. We demonstrate the framework on a lamella prepared by conventional FIB lift-out extracted from a semiconductor chip on a standard single-tilt holder: seven simultaneous HAADF+EDX projections within ±30° yielded a segmented 3D chemical volume consistent with a reference reconstruction from 49 projections over ±60°, with the complete workflow completing in under two hours. The contribution is methodological: a reconstruction framework that removes the geometric, computational, and operator burdens which have kept multimodal electron tomography out of routine FA practice.