Generating multiple-folding-geometry freeform reflective imaging systems based on optical design model-informed neural network

Freeform systems play important roles in modern optical systems, but their design remains challenging due to the complexity of freeform surfaces and lack of efficient methods as well as reference designs. This paper presents a framework that leverages an optical design model-informed neural network (ODMINN) to automatically generate multiple-folding-geometry freeform reflective imaging systems. The network is trained by both the data-driven loss and the physics-informed loss. An automatic training dataset generation method, combined with a fast light-obstruction evaluation method based on equivalent spherical systems, is proposed for obtaining dataset containing systems with various parameters and folding geometries. The real optical design model is integrated into the training process, by directly calculating the physics-informed loss related to imaging performance and optical design constraints using differential ray tracing. Freeform systems can be generated immediately by the network based on the design requirements. Compared with previous network which can only generate systems with one specific folding geometry, multiple-folding-geometry freeform systems can be generated using the proposed framework. We demonstrate the framework by designing freeform off-axis three-mirror systems with all eight different folding geometries. Our approach can significantly reduce human involvement and dependency on existing reference systems in the design of freeform optics, while dramatically improving the design efficiency.