Abstract
Nanoscale wafer surface variations degrade depth of focus and overlay accuracy in lithography, requiring precise surface topography measurement in semiconductor manufacturing. Interferograms encode phase variations induced by surface profiles, from which phase retrieval reconstructs the underlying topography. Phase retrieval from a single interferogram is an ill-posed inverse problem, where accurate phase retrieval is challenged by complex fringe variations and insufficient exploitation of inherent fringe information. To address this challenge, a physics-informed phase retrieval method is proposed based on an adaptive wavelet transform for physics-consistent phase inference. An auxiliary branch learns a spatially varying parameterization of the wavelet transform, enabling adaptive wavelet analysis and the extraction of wavelet-derived representations. These representations are integrated into the phase retrieval network through cross-attention to provide physics-informed guidance. Experiments demonstrate consistent improvements in phase retrieval accuracy on synthetic data and competitive performance on real data, while evaluations on synthetic data indicate generalization under varying fringe patterns and interferogram degradations.