Breaking human priors: reinforcement learning for lens design

Expert systems and deep learning methods have been used to automate starting-point lens designs and assist optimisation. While effective for well-defined tasks, these approaches depend on curated databases and human expertise, and typically require pre-defining optical element sequences. Such priors limit exploration of the design space and constrain what the model can learn, suppressing discovery of unconventional or optimal configurations, particularly for complex refractive systems. To enable exploration of potential designs, we employ reinforcement learning (RL) guided only by Snell’s law to autonomously design refractive lenses. The agent rediscovers known strategies while exploring various configurations, uncovering optimal arrangements of lenses and air gaps.