High-precision misalignment sensing for lithography using a joint space-frequency regression network.

In the realm of lithography misalignment sensing, achieving sub-0.2 nm precision is fraught with significant obstacles, largely due to the exacting demands placed on regression algorithms tasked with analyzing moiré fringes. These challenges stem from the inherent limitations of convolutional networks that process information from only one domain, often resulting in a deficit of critical feature data necessary for high-precision regression. To overcome these hurdles, we have engineered a groundbreaking convolutional regression network that synthesizes both spatial and frequency domain information from fringe patterns. This holistic approach has successfully recorded misalignment measurements with remarkable accuracy, achieving 0.12 nm at a 3σ confidence level. The method has shown considerable robustness against both system errors and environmental noise, bolstering its suitability for critical applications.