Small-area marks realize nanoscale lithography alignment by a spatial and frequency domain fusion neural network

For traditional moiré-based lithography alignment technology, which is widely used in proximity lithography systems, complex alignment marks with larger areas are employed to achieve high-precision misalignment detection. However, every inch of space on the wafer is extremely precious in practice, leaving minimal space for alignment marks. Therefore, employing small-area alignment marks in lithography systems will be a very challenging task with considerable potential in the future. The primary challenge is that existing frequency-based analytical algorithms struggle to achieve misalignment values with high-precision from moiré fringe images generated by small-area marks. To address this challenge, a spatial and frequency information fusion neural network (SFFN) is proposed for processing the moiré fringe images. With SFFN, the area of the alignment mark can be reduced by 2/3, and the average error of SFFN is less than 1 nm on the test dataset.