Deep reinforcement learning-based lens design

Area of Science:

• Optical Engineering
• Artificial Intelligence
• Computational Optics

Background:

• Lens design is a complex, iterative process demanding significant expertise.
• Current AI methods for lens design often depend on existing designs, limiting innovation.
• Exploring novel AI-driven approaches for data-free lens design is crucial.

Purpose of the Study:

• To investigate the capability of an AI model using only Snell's law for generating sensible lens designs.
• To develop and evaluate a deep reinforcement learning (RL) framework for automated refractive lens design.
• To establish a baseline for data-free, goal-driven lens design automation.

Main Methods:

• A deep reinforcement learning (RL) framework was developed to enable an AI agent to learn lens design.
• The framework incorporated a hybrid action space for flexible insertion of optical elements and air gaps.
• Snell's law was the sole optical principle guiding the AI's design generation process.

Main Results:

• The RL agent successfully generated a diverse range of refractive lens designs.
• Training criteria significantly impacted the AI's performance, including training time, inference speed, and design diversity.
• Generated designs exhibited structural similarities to established optical systems.
• The trained RL agent demonstrated adaptability by producing acceptable designs for untrained requirements.

Conclusions:

• Deep reinforcement learning, using Snell's law alone, can effectively automate refractive lens design.
• This data-free approach offers a new paradigm for exploring novel optical configurations.
• The developed framework serves as a foundation for future AI-driven lens design advancements.