Generating starting points for designing freeform imaging optical systems based on deep learning

Deep learning is an important aspect of artificial intelligence and has been applied successfully in many optics-related fields. This paper proposes a generalized framework for generation of starting points for freeform imaging optical design based on deep learning. Compared with our previous work, this framework can be used for highly nonrotationally symmetric freeform refractive, reflective, and catadioptric systems. The system parameters can be advanced and the ranges of these system parameters can be wide. Using a special system evolution method and a K-nearest neighbor method, a full dataset consisting of the primary and secondary parts can be generated automatically. The deep neural network can then be trained in a supervised manner and can be used to generate good starting points directly. The convenience and feasibility of the proposed framework are demonstrated by designing a freeform off-axis three-mirror imaging system, a freeform off-axis four-mirror afocal telescope, and a freeform prism for an augmented reality near-eye display. The design framework reduces the designer's time and effort significantly and their dependence on advanced design skills. The framework can also be integrated into optical design software and cloud servers for the convenience of more designers.