Optical Lithography Proximity Correction using Machine Learning and Richardson-Lucy Deconvolution

The proximity effect in optical lithography distorts patterns due to diffraction, scattering, and interference, reducing accuracy and increasing defects. This study combines Machine Learning and iterative deconvolution to enhance pattern precision. A dataset of mask patterns distorted by Gaussian, Sinc, and Airy disk Point Spread Functions trains a model to estimate their contributions. For a sixteen-square pattern with fifteen-micron sides and three-micron spacing, Gaussian blurring causes a 10% area spread prompting proportional decrease. The Richardson-Lucy Deconvolution algorithm refines the pattern by deconvolution, yielding squares with twelve-micron sides and six-micron spacing. Evaluations show improvements with a Peak Signal-to-Noise Ratio of 24.1940, Structural Similarity Index of 0.98644, Edge Normalized Cross-Correlation of 0.14976, and Deconvolution Quality factor of 0.99619. This method effectively mitigates proximity effects and enhances mask-writing accuracy.