Optimizing surface reconstruction using a segmented scanning electron microscope detector and deep learning

Accurate depth measurements of on-chip structures are increasingly important in semiconductor metrology. Given that Scanning Electron Microscopes (SEMs) are already widely used for fast, localized 2D imaging, exploring their potential for 3D capabilities is highly interesting. This paper presents an optimization study focused on 3D reconstruction performance when angle-resolved detection is achieved through a partitioned back-scattered electron (BSE) detector. The conducted research begins with an analysis of electron distribution projections on the detector surface, simulated using a Monte Carlo scattering model for relevant structures. Then a post-processing module is implemented which enables the simulation of various detector configurations that serve as input to an optimized machine-learning model. Using this framework, we evaluate and compare multiple detector setups, demonstrating a 56% depth prediction accuracy improvement over the single-detector baseline. Additionally, improvements ranging from 32% to 49% on typical critical dimension metrology are achieved using a multi-detector setup. This work highlights the benefits of a data-driven, angle-resolved detection approach to improve metrology precision for future semiconductor nodes.