OCT 23, 2026

Virtual metrology for improving overlay control capability in high-NA half-field exposure

SPIE Photomask Technology + EUV Lithography

Authors

Jihoon Jung, Jeng-Hun Suh, Taekwon Jee

Abstract

Sparse physical sampling can leave a substantial gap between achieved overlay residual and the nominal capability of high-order correction per exposure (CPE). We test whether ideal virtual observations can recover the correction-relevant information missing from a minimally redundant CPE33 fit in High-NA half-field exposure. A noise-free synthetic raw-overlay population comprised 100 independent 300-mm-wafer realizations, each with 131 complete 26 × 16.5 mm exposures and 51 reticle-compatible scribe-lane candidate positions. The baseline used N = 17 observations per exposure. Raw overlay was the pre-correction vector sum of wafer-scale smooth deformation, an exposure-common intrafield fingerprint, field-to-field variation, and high-spatial-frequency content. At equal added budgets M = 3, 6, and 9, fixed-position ideal observations were compared with dense-truth-informed greedy observations selected from the same 34 unobserved candidates. Ideal values entered only the CPE fit; no virtual-metrology predictor was implemented. Median dense-domain P99 residual was 6.25 nm for the baseline and 1.77 nm for a dense-fit CPE33 reference. Fixed-position addition reduced P99 to 3.98, 2.98, and 2.80 nm, showing that untargeted information recovered part, but not all, of the capability gap. Targeted selection reached 2.84, 2.30, and 2.13 nm and provided paired reductions of 0.84 [0.66, 1.02], 0.78 [0.61, 0.90], and 0.73 [0.51, 0.86] nm beyond fixed-position addition. At M = 9, the targeted benchmark closed approximately 92% of the aggregate-median baseline-to-dense-fit gap. Thus, the upper-bound value of VM in this experiment is targeted recovery of correction-relevant information that sparse sampling misses. This is a privileged-information benchmark, not demonstrated VM prediction performance.