Abstract
A system and method are disclosed for adaptive selection of a measurement region-of-interest (ROI) in metrology images, such as semiconductor overlay and critical dimension metrology. Instead of a binary inclusion/exclusion mask, a soft ROI is used: a pixel-wise weight map assigning each pixel a continuous value representing its reliability and informational contribution, so that pixels dominated by edge diffraction, stray light or neighboring-structure interference are down-weighted rather than discarded. The weight map is produced by a machine-learning model, in one embodiment a Physics-Informed Neural Network trained with a composite objective comprising a measurement accuracy term based on reference metrology data (e.g., CD-SEM), a physical consistency term derived from electromagnetic simulation (e.g., RCWA or FEM) of pixel-wise sensitivity, and a spatial smoothness term. Training can begin from simulation before real data exists, and physical consistency provides anomaly detection. Further embodiments include error-guided ROI optimization for trueness, defocus-aware and inter-layer differential ROI, frequency-domain ROI selection for small targets, real-time per-target adaptation, uncertainty-driven partial re-measurement, federated learning across a fleet of metrology tools, cross-modality data fusion, and joint optimization of the weight map and measurement recipe. Disclosed in South African provisional patent application No. 2026/04001, filed 14 April 2026. Parties interested in collaboration, licensing or development during the priority period are invited to contact the author at [email protected].
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
ARIELI, YOEL SHIMO, "ADAPTIVE SOFT REGION-OF-INTEREST SELECTION FOR METROLOGY USING PHYSICS-INFORMED MACHINE LEARNING", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12358