Deep reinforcement learning-based lens design.

Lens design is an important and time-consuming process that requires expert knowledge and multiple iterations of parameter adjustment and optimization to meet the required optical performance. While artificial intelligence (AI)-based approaches using expert systems and deep learning with curated databases have shown promise in automating lens design, their dependence on existing designs and pre-defined sequences of optical elements inherently restricts the exploration of new designs within the solution space. In this paper, we aim to investigate whether an AI-based model with Snell's law alone is capable of producing sensible lens designs. We present a deep reinforcement learning (RL)-based framework that enables an agent to learn to generate refractive lens designs. The proposed framework features a hybrid action space, allowing the agent to flexibly insert glass elements or air gaps, along with their parameters, into the optical layout. Evaluation results demonstrate the capacity of RL to generate a diverse range of lens designs, and highlight the impact of different training criteria on the training process, inference speed, reward design, and design diversity. We also examine the structural characteristics of the generated designs and find similarities with well-established designs in the literature. Furthermore, the trained RL agent is evaluated under untrained design requirements, where it successfully produces acceptable optical configurations. This work provides both a foundation for data-free, goal-driven lens design automation and a baseline for future RL-based approaches in lens design.