Reinforcement learning enabled the design of compact and efficient integrated photonic devices Mirbek Turduev1 Emre Bor2 Onur Alparslan3 Y. Sinan Hanay4 Hamza Kurt5

2025-04-29 0 0 669.9KB 6 页 10玖币

侵权投诉

Reinforcement learning enabled the design of

compact and efficient integrated photonic devices

Mirbek Turduev1, Emre Bor2, Onur Alparslan3, Y. Sinan Hanay4, Hamza Kurt5,

Shin’ichi Arakawa3, and Masayuki Murata3

1Department of Electrical and Electronics Engineering, Kyrgyz Turkish Manas University, Bishkek

720042, Kyrgyzstan

2Department of Electrical and Electronics Engineering, TOBB University of Economics and

Technology, Ankara 06560, Turkey

3Graduate School of Information Science and Technology, Osaka University, Osaka 565-0871, Japan

4Department of Computer Engineering, Akdeniz University, Antalya 07070, Turkey

5School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST),

Daejeon, 34141, Republic of Korea

E-mail: mirbek.turduev@manas.edu.kg

Received xxxxxx

Accepted for publication xxxxxx

Published xxxxxx

Abstract

In this paper, we introduce the design approach of integrated photonic devices by employing

reinforcement learning known as attractor selection. Here, we combined three-dimensional

finite-difference time-domain method with attractor selection algorithm, which is based on

artificial neural networks, to achieve ultra-compact and highly efficient photonic devices with

low crosstalk such as wavelength demultiplexers and a polarization splitter. The presented

devices consist of silicon-on-insulator materials, which are compatible with complementary

metal-oxide-semiconductor technology, and their structural dimensions enable the possible

fabrication process in the future. The numerical results are presented for the near-infrared

wavelengths at around 1550 nm, and the performance of designed photonic devices with

footprint of 33 µm2 are compared with the previously reported structures. Consequently, the

reinforcement learning is successfully applied to design smaller and superior integrated

photonic devices where the use of presented approach can be further expanded to different

applications.

Keywords: integrated photonic devices, wavelength demultiplexer, polarization beamsplitter, machine learning, artificial

neural networks, reinforcement learning

1. Introduction

Design of photonic devices are, in general, based on

physical knowledge and intuition where small number of

structural parameters are finely tuned by hand. Nevertheless,

brute-force designs are incompetent to reveal efficient

photonic devices. In the last decade, novel design approaches

introduced which mainly integrate optimization algorithms

with numerical methods to form remarkable photonic devices

by genuinely calculating the interaction between light and

matter. In this manner, various photonic devices are

presented such as subwavelength focusing lenses, mode

order converters, asymmetric light transmitters and optical

cloaks [1–6]. Evidently, these approaches took the design of

photonic devices one step further.

In recent years, the so-called “inverse-design” methods,

which use mostly silicon-on-insulator (SOI) materials, are

introduced to design ultra-compact, highly efficient and

complementary metal-oxide-semiconductor (CMOS)-

compatible integrated photonic devices. For instance, an

approach based on convex optimization is presented to

design several wavelength demultiplexers [7–10]. Also,

direct-binary search algorithm is applied to design a

polarization beamsplitter and a polarization rotator [11, 12].

Lately, machine learning and artificial neural networks

have attracted great interest from researchers and enabled

design of photonic devices in a different manner. For

example, deep learning accelerated the design of an all-

dielectric metasurface structures [13, 14]. Also, Bragg

grating devices are obtained by training artificial neural

networks [15]. In addition, a machine learning algorithm is

proposed for focusing and optical coupling of light [16].

Moreover, a reinforcement learning is applied to design

optical coupler and asymmetric light transmitter devices [17].

In the present study, we demonstrate the design of

wavelength demultiplexers and a polarization beamsplitter

operating at near-infrared wavelengths via attractor selection

(AttSel) algorithm which is considered as reinforcement

learning and based on artificial neural networks [18, 19]. For

this purpose, we combined the algorithm and three-

dimensional (3D) finite-difference time-domain (FDTD)

method [20]. The designed structures exhibit high optical

performance in an ultra-compact area and their numerical

investigations are presented in detail. It should be noted that

fabrication constraints are considered throughout the design

process [11] which enables the possible fabrication of

devices in the future. We believe that the proposed approach

is not only restricted to design integrated photonic devices,

but it may lead the advances in different photonic designs.

2. Design approach and numerical investigation of

integrated photonic devices

Even though photonic integrated circuits have superior

features comparing to integrated electronic circuits, they

have a main deficiency of lower integration density [21, 22].

On the other hand, there is a trade-off between optical

performance and structural dimensions of a photonic device.

For this reason, designing a highly efficient photonic

structure with small footprint is a challenge which can be

overcome by applying advanced search algorithms or

machine learning.

In general, machine learning is divided into two branches

such as supervised learning and unsupervised learning. In

these branches, simply, a data set is considered to train and

test the learning algorithm where data are independently

collected, i.e., the algorithm does not have any effect on

during the collection process of data. On the other hand,

artificial neural networks (ANNs) are considered as the

subclass of machine learning methods which is powerful for

modelling nonlinear relationships.

In a very recent study, ANNs are applied to characterize

the cases of “forward” and “inverse” designs [15]. An ANNs

is applied to map four structural parameters of Bragg

gratings to two optical characteristics. In forward modelling,

the selected four structural parameters are considered as

input to ANNs whereas the two optical characteristics are

used as output. In inverse modelling, the input and output

parameters are switched. In other words, two optical values

are inserted as inputs to ANNs whereas four structural values

are introduced as outputs. As a result, as expected, ANNs

easily found the nonlinear relationship between small

number of input and output parameters. In similar cases,

ANNs is applied for a regression problem which is

considered as a branch of supervised learning.

However, in the studies of inverse design of photonic

devices, hundreds of structural parameters are optimized to

find only one or small number of optical characteristics [1–

12]. In this case, for inverse modelling, a simple regression

method based on ANNs would not be able to find the

characteristics between large number of structural parameters

and small number of output values. The reason of this

possible failure is that the small number of input parameters

may not be informative enough for ANNs to map them to

large number of output parameters. Moreover, in the case of

inverse modelling, if a data set does not contain any results

of good optical characteristics, ANNs would not predict

structural parameters for a desired optical performance since

ANNs do not operate as search/optimization algorithms.

To overcome this issue, another branch of machine

learning known as reinforcement learning would be

reasonable to apply for photonic designs. In the

reinforcement learning, the algorithm works during the

sampling of data set which differs from supervised and

unsupervised learning. Therefore, a reinforcement learning

algorithm can find the values of large number of structural

parameters for even small number of desired optical

characteristics. For this reason, in our recent work, we have

applied AttSel algorithm which is a reinforcement learning

method, to design optical coupler and asymmetric light

transmitter [17].

AttSel models the interaction of the metabolic reaction

network and the gene regulatory network in a cell [18]. The

cell growth requires converting the nutrition in the

environment by metabolic reactions of proteins to the

substances necessary for the growth. The proteins that carry

out this conversion are produced by the gene reaction

network, where each gene has an expression level for

controlling the protein production level. The expression

levels change with the rate of substance production. If the

substance production rate is high, it implies that the

conditions are favorable, so the state is saved as an attractor

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ReinforcementlearningenabledthedesignofcompactandefficientintegratedphotonicdevicesMirbekTurduev1,EmreBor2,OnurAlparslan3,Y.SinanHanay4,HamzaKurt5,Shin’ichiArakawa3,andMasayukiMurata31DepartmentofElectricalandElectronicsEngineering,KyrgyzTurkishManasUniversity,Bishkek720042,Kyrgyzstan2DepartmentofEl...

收起<<

Reinforcement learning enabled the design of compact and efficient integrated photonic devices Mirbek Turduev1 Emre Bor2 Onur Alparslan3 Y. Sinan Hanay4 Hamza Kurt5.pdf

共6页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Reinforcement learning enabled the design of compact and efficient integrated photonic devices Mirbek Turduev1 Emre Bor2 Onur Alparslan3 Y. Sinan Hanay4 Hamza Kurt5

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: