A physical-model-based multi-level deconvolution network to achieve nanoscale measurement accuracy for optical critical dimension metrology

Abstract This study proposes a physical-model-based multi-level deconvolution network (PMLDNet) to enhance the lateral resolution of optical microscopy imaging for optical critical dimension (OCD) metrology. Unlike most AI-driven traditional methods, the PMLDNet uses Fourier optics to define the point spread function (PSF), which is essential for high-resolution microscale imaging. PMLDNet integrates spatially-variant deconvolution with a PSF grid, employing multiple kernels to extract features from microscopic images and PSF data. The PMLDNet models nonlinear interactions and, combined with an extensive PSF database, surpasses traditional resolution limits and eliminates ringing artifacts common in standard Wiener deconvolution techniques. This method enables precise measurement of sub-micron grating critical dimensions (CD), achieving measurement biases of less than 10 nanometers in spacing and linewidth CD measurement with a standard deviation of less than 5 nanometers for micron-scale structures. The developed approach improves CD measurement accuracy by 70% and offers a 57% enhancement over the classical optimized Wiener method, making it highly beneficial for OCD metrology and automated optical inspection in semiconductor manufacturing.