Top-profile-prior-constrained inversion for bottom-parameter reconstruction of extreme ultraviolet mask phase defects

Accurate repair or compensation of buried multilayer phase defects in extreme ultraviolet (EUV) lithography requires reliable knowledge of their three-dimensional morphology, yet the bottom-profile parameters remain difficult to recover from a single aerial-image observation. Through controlled diagnostic simulations, we show that this limitation originates primarily from the observation rather than from insufficient network capacity: the aerial image is much more sensitive to the defect top profile than to the buried bottom profile, and the error therefore saturates even when the network is trained only on the bottom parameters. We address this information bottleneck by reformulating reconstruction as a conditional inverse problem in which the independently measurable top profile, physically coupled to the bottom through multilayer deposition smoothing, is used as an explicit prior. The resulting dual-stream network, CoPriNet, injects the prior into hierarchical image features through latent-conditioned feature-wise linear modulation and enforces the deposition-smoothing inequality by construction. On simulated EUV mask phase defects, CoPriNet reduces the bottom-parameter mean relative error to 0.91% for bump defects and 0.79% for pit defects, while remaining robust to noise in the reconstructed complex field and to moderate top-prior error.