Related Experiment Video
Updated: Jun 5, 2025
08:17
Free-form Light Actuators — Fabrication and Control of Actuation in Microscopic Scale
Published on: May 25, 2016
9.2K
Sample-efficient inverse design of freeform nanophotonic devices with physics-informed reinforcement learning
Chaejin Park1,2, Sanmun Kim1, Anthony W Jung2
1School of Electrical Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea.
Nanophotonics (Berlin, Germany)
|December 16, 2024
Summary
Physics-informed reinforcement learning (PIRL) accelerates nanophotonic device design by significantly improving sample efficiency and overcoming local minima. This method achieves record-breaking performance in metasurface beam deflectors and offers robust optimization for complex freeform devices.
Area of Science:
- Nanophotonics
- Computational Physics
- Machine Learning
Background:
- Designing optimal nanophotonic devices is challenging due to vast combinatorial spaces.
- Conventional optimization methods often suffer from low sample efficiency and local minima.
Purpose of the Study:
- To introduce a novel optimization method, physics-informed reinforcement learning (PIRL), for freeform nanophotonic design.
- To demonstrate PIRL's superior sample efficiency and ability to avoid local minima compared to conventional methods.
Main Methods:
- Combining adjoint-based methods with reinforcement learning (RL).
- Developing reward engineering strategies to incorporate design constraints, such as minimum feature size.
- Utilizing transfer learning to further enhance sample efficiency.
Main Results:
- Achieved an order of magnitude improvement in sample efficiency over conventional RL.
- Designed high-performance one-dimensional metasurface beam deflectors, surpassing most reported records.
- Demonstrated robustness and the ability to integrate practical design constraints.
Conclusions:
- PIRL offers a highly sample-efficient and robust approach for optimizing complex freeform nanophotonic devices.
- The method shows promise for accelerating discovery in various physical domains requiring combinatorial optimization.
