CAUNet-MPC: Contour-aware U-Net for efficient and high-fidelity self-supervised mask process correction

In photomask manufacturing, process-induced contour errors after electron-beam writing and downstream processing can limit pattern fidelity even when the target layout is geometrically well defined. Mask process correction (MPC) precompensates the written layout, but conventional model-based MPC can require repeated process evaluation and local geometry updates for dense, contour-rich layouts. This study presents CAUNet (Contour-Aware U-Net)-MPC, a learning-assisted workflow that combines supervised process identification and reference-mask initialization with self-supervised process-consistency optimization. Under one fixed industrial electron-beam lithography condition, contour-aware U-Nets are trained from coordinate-aligned target patterns and measured contours and then fixed as differentiable process surrogates. An inverse CAUNet initialized from reference MPC masks is subsequently trained by a cross-surrogate minimax objective over two heterogeneous frozen process models. On a guarded spatial test split, three process surrogates achieved mean intersection-over-union (IoU) values of 0.9895–0.9899 and mean tile EPE95 values of 1.937–1.981 nm against measured contours. Under PA, CAUNet-MPC reduced mean tile EPE95 from 33.288 nm for No MPC, 13.184 nm for the actual industrial reference masks, and 13.056 nm for the locked warm-start network to 1.862 ± 0.062 nm across three single-network correction chains, while increasing IoU to 0.995 552 ± 0.000 222. The second optimization surrogate and a held-out architecture-transfer surrogate yielded 1.866 ± 0.205 and 1.877 ± 0.088 nm, respectively. These results establish CAUNet-MPC as a robust self-supervised route for fixed-process MPC candidate generation.