FreeformNet: fast and automatic generation of multiple-solution freeform imaging systems enabled by deep learning

Abstract

Using freeform optical surfaces in lens design can lead to much higher system specifications and performance while significantly reducing volume and weight. However, because of the complexity of freeform surfaces, freeform optical design using traditional methods requires extensive human effort and sufficient design experience, while other design methods have limitations in design efficiency, simplicity, and versatility. Deep learning can solve these issues by summarizing design knowledge and applying it to design tasks with different system and structure parameters. We propose a deep-learning framework for designing freeform imaging systems. We generate the data set automatically using a combined sequential and random system evolution method. We combine supervised learning and unsupervised learning to train the network so that it has good generalization ability for a wide range of system and structure parameter values. The generated network FreeformNet enables fast generation (less than 0.003 s per system) of multiple-solution systems after we input the design requirements, including the system and structure parameters. We can filter and sort solutions based on a given criterion and use them as good starting points for quick final optimization (several seconds for systems with small or moderate field-of-view in general). The proposed framework presents a revolutionary approach to the lens design of freeform or generalized imaging systems, thus significantly reducing the time and effort expended on optical design.