Main menu



Artificiel Inteligence Free Structure Metasurface Optimization

Artificiel Inteligence (AI) Free Structure Metasurface Optimization

KAIST announced on the 25th that Professor Min-seok Chang's research team in the Department of Electrical and Electronic Engineering had proposed a free-structure meta-surface structure design method based on reinforcement learning through joint research with Dr. Chan-yeon Park of KC ML2 (a research organization established by KC, a semiconductor manufacturing solution company).

Introduction about Artificiel Inteligence (AI): reinforcement learning.

Metasurface refers to a nano-optical device that achieves unprecedented properties of light using a structure much smaller than the wavelength of light. Nano-optical devices control the characteristics of light at the micro level, and can be used for LiDAR beam steering devices used for autonomous driving, ultra-high-resolution imaging technology, optical properties control of light emitting devices used in displays, and hologram generation. . Recently, as the expected performance of a nano-optical device increases, interest in optimizing a device having a free structure in order to achieve a performance far exceeding that of the device structure in the past is increasing. This is the first case of solving a problem with a large design space such as a free structure by applying reinforcement learning.


The Artificiel Inteligence (AI) research:

This research, in which KAIST Researcher Seo Dong-jin and ML2 Researcher Nam Won-tae participated as co-first authors, was published as the front cover thesis of the February 2022 issue of the international journal 'ACS Photonics'. (Paper Title: Structural Optimization of a One-Dimensional Freeform Metagrating Deflector via Deep Reinforcement Learning)

Reinforcement learning is an artificial intelligence methodology that mimics how animals learn. In animal behavioral psychology, an experiment known as 'Skinner's box' was the motif. At first, the rat, which was acting randomly, presses the lever and confirms that the food comes out. Over time, the lever is pressed with a higher frequency, and it can be observed that a certain reward (prey) 'reinforces' the behavior (the act of pressing the lever). Reinforcement learning, which has a structure very similar to the experiment, is an artificial intelligence methodology in which a behavioral subject learns about the environment while receiving a 'reward' from the 'environment' surrounding him.

AlphaGo example:

Google DeepMind's 'AlphaGo', which won the match against Lee Sedol 9-dan in 2016, is a representative example. AlphaGo learned the complex rules of Go through interaction with the environment represented by the checkerboard, and was able to make a near-optimal choice among known cases where there are more atoms in the universe. Recently, reinforcement learning in the artificial intelligence academia has been receiving a lot of attention as the most similar form of artificial intelligence to human intelligence.

Interpretation of research team about Artificiel Inteligence (AI): reinforcement learning.

The research team proposed an idea to utilize the features of reinforcement learning, which can easily learn complex environments, for optimization of meta-surface free structures. Previously, the metasurface free structure optimization technique was considered to be difficult to solve due to the number of too many cases. Therefore, the existing research direction mainly utilized a method of simplifying the structure with simple basic figures. However, this method had limitations in that the geometric structure was limited, and optimization techniques for more complex structures were considered difficult to achieve.

The algorithm proposed by the research team starts with a very simple idea. The 'behavior' of reinforcement learning is defined as 'overturning' the components of a structure one by one. This was to overturn the idea of ​​optimizing free structures, which was previously thought only as a way to create structures as a whole. The research team showed that using this method, it is possible to broadly explore possible structures and find optimal structures without special prior knowledge of metasurfaces. In addition, efficiencies close to 100% have been achieved under many incident conditions, comparable to or better than state-of-the-art performance, and under certain conditions.

It is expected that this study will find a new breakthrough in the field of free structure optimization, and it is expected that it can be utilized not only for optical devices but also for device structure optimization in many fields.

Researcher Dongjin Seo, the first author, said, “Reinforcement learning is an effective algorithm for finding the optimal case in a complex environment. In this study, we are happy to leave a successful case of performing optimization of free structures with this method.”

KAIST Professor Min-seok Chang said, “I hope that good results will come out in the field of applying artificial intelligence technology to optical engineering and contribute to raising the status of science.”

Meanwhile, this research was conducted with support from the National Research Foundation of Korea's Mid-Range Researcher Support Project (Strategic Research), Korea-Switzerland Innovation Program, and Future Materials Discovery Project.