An Integrated Physics-Based and Data-Driven Framework for Defect Prediction in Advanced Nanoimprint Lithography Toward Inorganic Semiconductor Patterning

Advanced nanoimprint lithography (NIL) is promising for inorganic semiconductor patterning because it enables high-resolution replication with a relatively simple process flow; however, yield loss increasingly originates from spatially distributed, subcritical distortions accumulated across coating, exposure, etching, and imprinting. In this study, we propose an integrated physics-based and data-driven framework for pre-manufacturing defect-risk prediction in NIL. The framework combines an NDA-safe layout database, a physics-based process twin, and a stochastic risk prediction model using a physics-augmented convolutional neural network with conformal uncertainty calibration. Starting from binary design layouts, the process twin sequentially captures resist thickness variations during spin coating, proximity-induced dose redistribution and development-induced pattern deformation during electron-beam lithography (EBL), density-sensitive pattern transfer during reactive ion etching (RIE), and three-dimensional resist filling during imprinting, thereby generating physically consistent parameter maps for downstream learning. The results demonstrate an end-to-end virtual inspection flow that converts layouts into spatially resolved risk maps before fabrication. In addition, patterns with similar contour extent but different local density exhibit distinctly different risk distributions, indicating that manufacturability is governed not only by nominal geometry but also by local pattern environment. These findings support pre-manufacturing virtual inspection as a physically interpretable route for early yield-risk screening in advanced NIL.