Multiple-folding-geometry freeform reflective afocal systems design enabled by deep-learning-based neural network

Freeform off-axis reflective afocal systems show great potential in system compactness, structure flexibility, no light-obscuration and chromatic aberration. In this paper, we propose a framework for the generation of multiple-folding-geometry freeform off-axis reflective afocal systems based on an optical design model informed neural network. The data-driven loss is established for the initial training, and the physics-informed loss considering imaging performance and design constraints is obtained through differential ray tracing. The network is then trained jointly by combining the two losses. Once trained, the network can immediately generate one or more freeform afocal systems after inputting design requirements, which can be taken as good starting points for further fast optimization. Our method significantly improves design efficiency and greatly reduces reliance on complex design techniques.