Deep learning-based mask design method for film thickness uniformity in spherical rotation systems

Abstract The uniformity of the film thickness of large-aperture mirror is a critical factor affecting the imaging quality of reflective optical systems. A deep learning-based mask design strategy is proposed to reduce this non-uniformity. By developing a convolutional neural network architecture and generating training and validation datasets based on engineering experience, the model can produce a mask that fulfills application requirements following the optimization of network parameters. In comparison to conventional methods, this approach not only markedly decreases the number of experiments but also demonstrates considerable versatility. The trained neural network architecture can be utilized for various vacuum chamber configurations or optical components with diverse surface types, facilitating expedited mask design by merely supplying the relevant dataset. Theoretical results demonstrate that the mask designed using this technology diminishes the film thickness non-uniformity of a large-aperture mirror with a diameter of 3270 mm from 10.56% to 1.7%, so effectively validating its feasibility and superiority.