Accelerating 3D Semiconductor Chemistry Mapping by Automated TEM Tomography
Accelerating 3D Semiconductor Chemistry Mapping by Automated TEM Tomography
Thursday, October 8, 2026: 9:20 AM
Summary:
Multimodal electron tomography can produce 3D chemical maps of semiconductor devices at nanometre resolution but remains confined to research environments. Conventional workflows require needle specimens, dedicated tomography holders, tilt ranges of ±70° or more, 50–180 projections per modality, and manual multi-step reconstruction spanning a full day. We present a physics+AI reconstruction engine that solves 3D chemical tomography as a single self-supervised inverse problem. A physical forward model of the S/TEM-EDX imaging process is coupled with a neural network that parameterises the 3D chemical volume in a low-dimensional nonlinear space. The engine jointly recovers the volume and all nuisance parameters directly from raw data, with no pre-processing, no prior alignment, and no hyperparameter tuning. The engine was validated on 22 nm semiconductor structures prepared as conventional FIB lamellae on standard single-tilt holders. Only 7 correlative HAADF+EDX projections were acquired within ±30° in approximately 1 hour. Reconstructions showed good agreement with reference volumes from 49 projections over ±60°. The complete workflow completed in under two hours using only equipment already present in standard FA laboratories.
Multimodal electron tomography can produce 3D chemical maps of semiconductor devices at nanometre resolution but remains confined to research environments. Conventional workflows require needle specimens, dedicated tomography holders, tilt ranges of ±70° or more, 50–180 projections per modality, and manual multi-step reconstruction spanning a full day. We present a physics+AI reconstruction engine that solves 3D chemical tomography as a single self-supervised inverse problem. A physical forward model of the S/TEM-EDX imaging process is coupled with a neural network that parameterises the 3D chemical volume in a low-dimensional nonlinear space. The engine jointly recovers the volume and all nuisance parameters directly from raw data, with no pre-processing, no prior alignment, and no hyperparameter tuning. The engine was validated on 22 nm semiconductor structures prepared as conventional FIB lamellae on standard single-tilt holders. Only 7 correlative HAADF+EDX projections were acquired within ±30° in approximately 1 hour. Reconstructions showed good agreement with reference volumes from 49 projections over ±60°. The complete workflow completed in under two hours using only equipment already present in standard FA laboratories.