1

Role of Digital Twin in Optical Communication:

Fault Management, Hardware Configuration,
and Transmission Simulation
Danshi Wang, Zhiguo Zhang, Min Zhang, Meixia Fu, Jin Li, Shanyong Cai, Chunyu Zhang, and Xue
Chen

management, of optical communication systems and networks is

Abstract—Optical communication is developing rapidly in the significant and indispensable. However, with the proliferation of the
directions of hardware resource diversification, transmission elastic optical transceivers, complex optical modulation, hybrid optical
system flexibility, and network function virtualization. Its transmission, variable wavelength grids, software-defined optical
proliferation poses a significant challenge to traditional optical
communication management and control systems. Digital twin
networks, and other technologies, optical communication is rapidly
(DT), a technology that utilizes data, models, and algorithms and evolving in the direction of hardware resource diversification,
integrates multiple disciplines, acts as a bridge between the real transmission system flexibility, and network function virtualization [5,
and virtual worlds for comprehensive connectivity. In the digital 6]. Moreover, owing to the complex deployment environments and
space, virtual models are established dynamically to simulate and wide range of coverage, the existing optical communication systems
describe the states, behaviors, and rules of physical objects in the and networks encounter severe challenges in their accurate monitoring,
physical space. DT has been significantly developed and widely
applied in industrial and military fields. This study introduces the
on-line management, real-time control, and rapid maintenance [7]. In
DT technology to optical communication through addition, the conventional network management schemes always take
interdisciplinary crossing and proposes a DT framework suitable no account of actual physical situation. In researches of optical
for optical communication. The intelligent fault management network, there is a premise that the conditions of physical layer are
model, flexible hardware configuration model, and dynamic ideal, stable, and static, and the characteristics of physical layer are
transmission simulation model are established in the digital space mainly based on the simplified numerical models, but always ignore
with the help of deep learning algorithms to ensure the high-
reliability operation and high-efficiency management of optical
the real physical world. Therefore, there is a yawning gap between
communication systems and networks. network layer and physical layer. Although DT has been used widely
in various fields, it has not yet been applied to optical communication.
Index Terms—Digital twin, deep learning, optical In view of the technical challenges faced in optical communication,
communication. DT has the potential to provide low-cost and lightweight solutions to
address these challenges, and it may have broad application prospects
in optical communication.
INTRODUCTION Considering the need to develop optical communication, and
encouraged by the technical advantages of DT, this study introduces
C urrently, the world is entering the digital era, where digital
twins (DTs) have become an anticipated enabling
technology to promote digital transformation and intelligent Data
Fusion
Information
Mining
evolution in various areas [1]. DT, which creates high-fidelity
Database
digital virtual models of physical objects to simulate their
behaviors and depict their operating states, paves the way for
DIGITAL TWIN
realizing cyber–physical fusion [2]. The power of advanced
Physical Space

Digital Space

computing and analytics in the cyber space opens a bright Mapping

perspective to all walks of life. DT is receiving increasing Feedback
Physical object
attention and has been used in a wide range of fields, including Virtual model

the manufacturing industry, military engineering, marine Application

engineering, aerospace engineering, electrical power systems,

and smart cities [3, 4].
Owing to its high bandwidth, low loss, and strong anti-interference,
optical fiber communication is the key technology to realize large-
Smart
capacity and long-haul data transmission in this era of big data. Smart
workshop
Electrical
power
Ship
marine
Optical
communication
Aeronautical
engineering
Industry
manufacture city
Ensuring a stable operation, as well as efficient and intelligent Fig. 1. The basic principle and application background of digital twin.
 2

Digital twin for optical communication

Fault management Hardware config. Transmission simul.

Fault prediction Modulation format Dynamic simul.
Application Fault diagnosis Symbol rate Real-time display
layer
Deep Learning throughout the whole process Fault location FEC coding Numerical analyze

Strategy feedback & Iterative optimization

Fault repair Constel. shaping Damage comp.

Output response Model cooperative operating

Model Fault management Hardware config. Transmission simul.

layer model model model

Model establishment Data fusion processing

Data from Data from Data from

physical equipment transmission system network interface
Data
layer Physical character. Signal quality Network status
Operating state System perform. Network alarm

Data acquisition Sensing Monitoring Reporting

Physical
Optical communication system
layer
Fig. 2. Digital twin framework for optical communication consisting of physical layer, data layer, model layer, and application layer.

DT to optical communication to perform multiple functions, including industry, and smart cities (see Fig. 1). It has implemented the
fault management, hardware configuration, and transmission various services, such as intelligent manufacturing, work flow
simulation. First, a DT framework is designed particularly for optical control, operation status prediction, lifetime monitoring, system
communication. In the physical space, the acquisition and fusion of simulation, numerical analysis, 3D visualization, etc.
multi-source heterogeneous data from physical objects are Next, we carried out exploratory research on combination of
implemented using a cognitive optical network. In the digital space, an DT and optical communication. Firstly, the framework of DT
intelligent fault management model, a flexible hardware configuration particularly designed for optical communication is proposed,
model, and a dynamic transmission simulation model are established. which is composed of physical layer, data layer, model layer,
Deep learning as a powerful interdisciplinary tool plays a central role and application layer, as shown in Fig. 2. From the bottom up,
the physical layer in this architecture refers to the physical
in the entire DT system to realize all types of functions, including data
objects in optical communication system, i.e., optical
acquisition, model building, result analysis, and strategy execution.
equipment, network elements, transmission modules, fiber
The application of the DT technology in optical communication is
links, etc. Based on the idea of cognitive optical network,
realized to ensure a safe and stable operation of optical communication through sensor detection, optical performance monitoring, and
systems, optimize the hardware efficiency and network resource network message reporting, all sorts of data can be collected in
efficiency, and improve the dynamic monitoring and numerical real time from physical equipment, transmission systems, and
analysis capability of the transmission system. network interface sources. Then the acquired data are sent and
stored in the data layer, where the fusion processing and deep
DIGITAL TWIN FRAMEWORK FOR OPTICAL COMMUNICATION mining of multi-source heterogeneous data are carried out. Both
The basic principle and application background of DT are the unified data after fusion processing and the useful
depicted in Fig. 1. In the physical space, real-time data information obtained from data mining are subsequently
acquisition from physical objects is achieved through sensing transmitted to the model layer.
and monitoring. The collected data are processed and stored in Here, we mainly focus on the three essential subjects in
a database. After data mining and fusion processing, useful optical communication: fault management, hardware
information is obtained and deep-level knowledge is explored. configuration, and transmission simulation. Therefore, in the
Based on the collected data and captured information, virtual model layer, we investigate and build the intelligent equipment
models can be built dynamically in the digital space. According fault management model, flexible hardware configuration
to various application requirements, various virtual models model, and dynamic transmission process simulation model
generate the corresponding target results and feedback the simultaneously. Through the cooperative operation and
matching optimization strategies to the physical space. The dynamic response, these three models can perform the
reality–virtual interaction is realized by forward mapping to corresponding functions in the application layer, including the
virtual models and backward reaction to physical objects. Based equipment fault prediction/diagnosis/location/repairment,
on the idea of DT, a series of successful applications has been hardware resource optimization and network efficiency
demonstrated in the fields of smart workshops, electric power, improvement, transmission process dynamic simulation, real-
ship navigation, aeronautical engineering, manufacturing time display, and numerical analysis. Finally, according to the
 3

Data fusion processing

Data Data Data Data
balancing augmentation sampling normalization

Data Data Data Data Data

base desensitization cleaning labeling naming

Data layer

OTN equipment
WiFi access
Satellite access
Data
base

Edge cloud

Access node Convergence node Core node

2G/3G/4G/5G access Core cloud POT
XDSL/cable access
Transceiver
Physical layer of cognitive optical network
Fig. 3. Multi-source heterogeneous data collected from cognitive optical network consisting of physical layer of cognitive optical network and data layer for data
fusion processing.

comprehensive analysis results, the optimization strategies are are obtained from optical performance monitoring (OPM)
formulated and response actions are taken in the application modules. In addition, the transmitted signal can also be obtained
layer, and then feedback to the physical layer from the top down. by photoelectric detector. However, different data have
Through reality-virtual interaction between forward mapping different structural characteristics, operation modes, storage
and backward reaction, the DT system for optical mechanisms, and matching algorithms. Therefore, it is essential
communication forms a closed loop. As a supporting to implement data fusion for all the multi-source heterogeneous
technology, deep learning algorithm runs through the whole data in the data layer. Data fusion refers to data desensitization,
process to implement data acquisition, model establishment, cleaning, labeling, naming, normalization, sampling,
numerical analysis, and strategy execution [8]. Eventually, a augmentation, and balancing, which convert the original data
framework of DT specialized for optical communication system into processable data and usable information, as shown in Fig.
is established. 3.
Based on the data type, multi-source heterogeneous data can
DATA ACQUISITION be broadly divided into three categories: image data, time-series
The realization of DT is based on the massive data pertaining data, and other structural data. All these categories must be
to all elements and over the entire lifetime. A virtual model can comprehensively fused and further processed using deep
be accurately built in the digital space only when sufficient data learning. Various deep learning algorithms are available for this
are collected from the physical space. In optical communication, purpose. In terms of the algorithm mechanism and data
right from the physical layer to the network layer, significant category, the convolutional neural network, recurrent neural
amounts of data are generated from a wide range of sources, network, deep neural network, and their variants can be used to
including historical data, initial data, and real-time update data process image data, time-series data, and structured data [10,
for the network status, equipment operating state, and 11]. In addition, depending on research objects and modeling
transmission system performance. All types of multi-source requirements, deep learning can be used to extract important
heterogenous data are acquired from cognitive optical network features and capture deep-level information, such as the
[9], which aims to introduce cognition on multiple planes (e.g., prediction of equipment operating parameters, estimation of the
access, aggregation, and core nodes) to perceive current optical signal quality, evaluation of the performance of the
network conditions, as shown in Fig. 3. The data on the network transmission system, and analysis of the network status. All
status (e.g., resource utilization information, notification these are of significance for the establishment of a model.
message, delay jittering, and blocking rate) are obtained from
the network log viewer. The data on the equipment operating MODEL ESTABLISHMENT
state (e.g., environment temperature, unusable time, DT establishes a quasi-real-time connection between the
input/output optical power, and laser bias current) are acquired physical and digital worlds through virtual models. It uses
from sensing detectors. The data on the performance of the modeling technology to describe the physical state, simulate the
transmission system (e.g., optical signal-to-noise ratio (OSNR), operation process, forecast change tendency, and optimize
modulation format, transmission speed, and system impairment) object performance. Therefore, it can be said that the model
 4

layer is the most central part of the DT system. However, the

traditional modeling methods in optical communication suffer F3 32673 Laser Bias Current
from the following disadvantages: (a) strong dependence on
expert experience and mathematical theories, (b) effective only
for the static and single scenario, and (c) inability to perform F2 17268 Environmental Temperature
real-time iterative optimization. While in digital twin, massive
real-time data are collected and stored in the data layer. A

Feature
F9 17222 Output Optical Power
wealth of information and a variety of time-varying
characteristics are contained in the various data. Different from
traditional model-driven methods, in our scheme, based on the F13 13235 Input Optical Power
idea of real-time data-driven modeling, three deep learning-
based models are proposed: one each for intelligent fault
management, flexible hardware configuration, and dynamic F6 12482 Laser Temperature Offset
transmission simulation.
0 10000 20000 30000 1

FAULT MANAGEMENT MODEL Feature importance score

Large-scale optical networks consist of plenty of optical Fig. 4. Top 5 feature importance in XGBoost feature scoring for fault diagnosis.
equipment. A failure of the network can cause serious
consequences, such as data loss, transmission interruption, and
performance deterioration. The traditional approach to the fault node, branches, and leaf nodes. The path from the root node to
management of the optical network equipment is based on the the leaf node can explain the information of a certain decision
built-in inspection software provided by the manufacturer to rule. Encouraged by the idea that XGBoost finds the best split
monitor some typical operating parameters. When the real-time through feature splitting continuously, it is possible to introduce
monitoring values reach the alarm threshold value, the fault XGBoost to interpret the cause of the fault and reveal the
management system automatically sends out an alarm. This important features for fault diagnosis. The greater the number
type of fault management is simple and straightforward, and its of times the feature node splits, the stronger the correlation
technical architecture is relatively mature. However, parameter between the feature node and fault. Based on the data from
monitoring and threshold settings are generally based on operating state records of existing OTN, there are total 15
experience and lack precision. Furthermore, the traditional fault features that may cause equipment fault. By decomposing the
management system is capable of only a broad fault alarm and internal structure of XGBoost and analyzing node
fault prediction and lacks an effective fault diagnosis function. characteristics, the 5 most remarkable features responsible for
Therefore, there is a need for not only improving the accuracy the fault are found, as shown in Fig. 4. The feature scoring
but also extending the function scope of fault management. dependent on split time can be calculated in XGBoost. In terms
With the help of DT, optical network equipment equipped of their importance grades, the possible fault causes are the laser
with sensing and monitoring modules can collect various bias current, environment temperature, output optical power,
operating state parameters in real time, such as the environment input optical power, and laser temperature offset in order, which
temperature, laser power, bias current, etc. Based on this can provide valuable reference information for fault repair.
historical operating state data obtained from the equipment in
the physical space, virtual fault prediction and fault diagnosis HARDWARE CONFIGURATION MODEL
models are generated in the digital space using two learning
algorithms. First, a bidirectional gated recurrent unit (BiGRU) In dynamic optical networks, programmable optical equipment
algorithm is used to predict the trend of the equipment state serves as one of the key equipment that can flexibly adjust the
parameters that can be regarded as temporal sequences. Next, configuration parameters to meet the diverse quality of service levels
the predicted state parameters are used to prognosticate whether and adapt to the time-varying network environment. In this study, we
the equipment will fail at a certain time. use the programmable optical transceiver (POT) to investigate the
At present, we have demonstrated the feasibility of this application of DT in the hardware configuration. As the core network
scheme on optical transport network (OTN) equipment equipment, POT can flexibly configure multi-dimensional parameters,
(ZXOME 8300) those are mainly deployed in the metropolitan such as the modulation format (MF), symbol rate (SR), forward error
area network [12]. The data from preceding thirty-six days are correction (FEC) coding, probabilistic constellation shaping, and
collected for model training. Based on the trained model, the subcarrier, depending on the real-time link state, optical signal quality,
operating state of the following seven days can be predicted. It and service transmission requirements for the purpose of flexible
can predict the occurrence of the fault at least one day in bandwidth resource allocation and automatic transmission capacity
advance, with an average accuracy of greater than 99% and a regulation. The ideal POT configuration model can learn the nonlinear
false rate of less than 0.9%. mapping relationship between the POT operating parameters and the
Next, extreme gradient boosting (XGBoost) is selected to corresponding transmission performances under various network
implement fault diagnosis [13]. XGBoost is an integrated conditions to provide the optimal configuration strategy quickly.
learning algorithm that adopts the classification and regression However, the classical POT configuration model is mainly dependent
tree (CART) as the base learner. Unlike traditional neural
on expert experience, lacks a feedback mechanism, and is incapable of
networks, the base learner of XGBoost is composed of a root
a dynamic response. Therefore, it fails to meet the requirements of
 5

DRL can dynamically adjust the parameters of the control model

depending on the environmental changes, ensure the service
transmission requirements, and improve the utilization of the spectrum.
In the Fig. 5, the spectrum occupation decreases with the normalized
traffic load and less spectrum consumption can be obtained in the
proposed DRL-based scheme, where the average spectrum
consumption is about 77.9% of those of the BSMCP-based scheme.
Meanwhile, the average network delay performance curves are almost
overlapped in two schemes where the average delay in the proposed
scheme is 0.7% larger than those in BSMCP-based scheme. We are of
the view that DT will open a new avenue for adaptive optical
component modeling and control of dynamic optical networks.

TRANSMISSION SIMULATION MODEL

Fig. 5. Comparisons between the proposed DRL-based scheme and the BSMCP-
based scheme in terms of the spectrum consumption and the network delay. Performing simulations of optical communication systems is
essential for system designs. It helps to simulate the actual operating
process and characterize the real environment numerically. However,
dynamic optical networks. Encouraged by the strong dynamic digital
the traditional simulation systems for optical communication are
modeling and intelligent feedback controlling characteristics of DT,
usually composed of a series of model blocks, and each model block
we propose a DT-enabled POT adaptive configuration scheme using
is established based on rigorous mathematical theory. Therefore,
deep reinforcement learning (DRL), which is oriented by low delay
traditional simulation systems are generally effective only for static
and high spectral efficiency.
and ideal transmission scenarios and are not valid for practical
Based on the operating principle of DT, we propose a self-evolving
transmission scenarios (where the link characteristics and signal
POT based on DRL [14]. In the physical space, POT provides
performance vary dynamically in real time) and the newly evolving
transmission bandwidth resources flexibly to satisfy the diverse quality
transmission scenarios whose analytical models have not yet been
of service levels between the source and destination nodes. Through
accurately proposed and mathematically verified. Using DT and deep
configuring multi-dimension parameters of POT, the spectrum
learning, we propose a data-driven dynamic modeling technique for
consumption and delay performance are affected accordingly. In the
simulating the process of optical transmission instead of block-based
data base, multidimensional monitoring data on POT collected from
system, as shown in Fig. 6. According to the characteristics of deep
the physical space are uploaded to the data fusion platform for data
learning, the model functions can be approximated by mapping
desensitization, cleaning, and normalization. The data fusion platform
independent to dependent variables corresponding to the input and
comprises multi-dimensional monitoring data and information,
output data. Deep learning could construct an approximate channel
including but not limited to OSNR and control actions from physical
model driven by source data and received data. In the physical space,
layer, transmission capability and network delay from network layer,
signal data before transmission and after reception are collected in real
and bandwidth and delay requirements from application layer. After
time. Deep learning-based dynamic response models are established
data fusion, the POT monitoring data are input to the digital space.
in the digital space and driven by the collected data without
After fusion processing, the POT monitoring data are sent to the
dependence on experience and mathematical theory.
digital space, where DRL is used to implement the dynamic digital
In the physical space, the optical transmission system consists of an
modeling and devise the smart control strategy for the POT
optical transmitter, optical channel, optical amplifier, and optical
configuration. In the DRL agent, the double deep Q-learning network
receiver. In the digital space, the transmission process simulation
consisting of two deep neural networks—modeling and evolving
presents the entire process of optical signal transmission from the
networks with similar structures—is selected. The objective of the
transmitter, through fiber links, after relay amplifications to the
modeling network is to reduce network delay and improve the
receiver. Therefore, the optical signal data from transmission and
utilization of the spectrum. Through the historical and real-time POT
reception contain the entire real-time information, such as the signal
control experience, the modeling network is set to determine the
quality, channel characteristics, and equipment performance, of the
uncertain relationship between the network state and transmission
dynamic transmission process. In [15], we have demonstrated a data-
performance utilization under various POT control actions. In contrast,
driven fiber channel modeling using deep learning to simulate the
the evolving network is used to learn real-time POT control experience
transmission process through a standard single-mode fiber. First,
and, accordingly, adjust the parameters of the modeling network to
optical signals from the optical transmitting module and receiving
adapt to the dynamic network environment. This scheme can achieve
modules are collected as the input and output data to train the deep
a balance between transmission quality and spectrum resource cost;
learning model. The input and output optical signals are composed of
that is, it can significantly improve the utilization of the spectrum and
amplitude and phase information, which are typical time-series data.
reduce the network delay while meeting the transmission requirements
Considering the correlation between the front and rear sequences, the
for various services in various physical link transmission states.
multilayer bidirectional long short-term memory (BiLSTM) is
The experiment and simulation results indicate that, when
selected as the deep learning algorithm because of its superior
compared with the classical POT based on brute-force search and
performance for temporal sequences. Based on the amplitude and
maximum capability provisioning (BSMCP), the POT based on DT-
phase information of the optical signals, two BiLSTM networks are
 6

Block-based optical communication system

Source Symbol Shaping
data mapping filter
Fiber Optical
Amplifier Filter Detector Filter
Digital Received
sampling data
Laser Modulator

. .
. .
. .
.
.
.

Data-driven channel modeling based on Deep Learning

Fig. 6. Data-driven channel modeling based on deep learning to perform transmission simulation instead of block-based optical communication system constructed
in a divide-and-conquer manner.

used to execute amplitude and phase waveform fitting. Based on the

real-time collected amplitude and phase data of optical signals, REFERENCES
BiLSTM can effectively learn the transfer function and realize [1] F. Tao, J. Cheng, Q. Qi, M. Zhang, H. Zhang, and F. Sui, “Digital twin-
waveform fitting through iterative optimization training to depict the driven product design, manufacturing and service with big data,” The
Inter. J. Adv. Manufac. Technol., vol. 21, no. 9, pp. 3563-3576, 2018.
complete transmission process of the optical signal from its [2] E. Negri, L. Fumagalli, and M. Macchi, “A Review of the Roles of Digital
transmission to its reception in the digital space. Through testing and Twin in CPS-Based Production Systems,” Procedia Manufacturing, vol.
verification, we have realized a 0–80 km optical channel simulation 11, pp. 939-948, 2017.
model for the intensity modulation and direct detection system [15], [3] E. J. Tuegel, A. R. Ingraffea, T. G. Eason, and S. M. Spottswood,
“Reengineering aircraft structural life prediction using a digital
which proves the feasibility of the data-driven modeling scheme. twin,” Intern. J. Aero. Eng.., pp. 1687-5966, 2011.
Furthermore, because the scheme does not rely on expert experience, [4] F. Tao, H. Zhang, A. Liu, and A. Y. Nee, “Digital Twin in Industry: State-
it can significantly reduce the modeling cost and improve the of-the-Art,” IEEE Trans. Indus. Inform., vol. 15, no.4, pp. 2405-2415,
2019.
simulation efficiency. This transmission simulation model in DT [5] Z. Zhu, W. Lu, L. Zhang, and N. Ansari, “Dynamic Service Provisioning
system can not only digitize the physical process, but also provide the in Elastic Optical Networks with Hybrid Single-/Multi-Path Routing,”
numerical channel model that is important for adaptive damage IEEE J. Lightw. Technol., vol. 31, no. 1, pp. 15-22, Jan. 2013.
[6] P. Lu, L. Zhang, X. Liu, J. Yao, and Z. Zhu, “Highly-Efficient Data
compensation, like end-to-end learning method, to ensure high reliable Migration and Backup for Big Data Applications in Elastic Optical Inter-
transmission of optical communication. Data-Center Networks,” IEEE Netw., vol. 29, no. 5, pp. 36-42, 2015.
[7] F. Musumeci et al., “An overview on application of machine learning
techniques in optical networks,” IEEE Commun. Surv. Tut., vol. 21, no. 2,
CONCLUSIONS pp. 13831408, 2018.
In this study, we investigated the role of DT technology in optical [8] F. N. Khan, Q. Fan, C. Lu, and A. P. T. Lau, “An Optical
Communication's Perspective on Machine Learning and Its Applications, ”
communication and discussed its application for fault management, J. Lightwave Technol., vol. 37, no. 2, pp. 493-516, 2019.
hardware configuration, and transmission simulation. A DT [9] I. De Miguel et al., “Cognitive dynamic optical networks,” IEEE/OSA J.
framework suitable for optical communication was designed, the Opt. Commun. and Netw., vol. 5, no. 10, pp. A107-A118, 2013.
[10] J. Li, D. Wang, S. Li, M. Zhang, C. Song, and X. Chen, “Deep learning
problem of mass data acquisition and fusion in the physical space was
based adaptive sequential data augmentation technique for the optical
analyzed, and multiple deep learning-based virtual models were network traffic synthesis,” Opt. Exp., vol. 27, no. 13, pp. 18831–18847,
proposed. DT technology can effectively improve and ensure the 2019.
control efficiency and stable operation of optical communication [11] L. Deng, “A tutorial survey of architectures, algorithms, and applications
for deep learning,” APSIPA Trans. Signal Inf. Process., vol. 3, no. 2, pp.
systems and promote their digital transformation and intelligent 1–29, Jan. 2014.
evolution. In addition to the three types of application scenarios [12] C. Zhang, D. Wang, L. Wang, J. Song, S. Liu, J. Li, L. Guan. Z. Liu, and
discussed in this study, DT can also be used to realize other potential M. Zhang, “Temporal data-driven failure prognostics using BiGRU for
optical networks,” IEEE/OSA J. Opt. Commun. and Netw., vol. 12, no. 8,
applications, such as the lifecycle management and 3D visualization pp. 227-287, 2020.
of optical equipment, adaptive damage compensation of transmission [13] C. Zhang, D. Wang, C. Song, L. Wang, J. Song, L. Guan. Z. Liu, and M.
systems, intelligent monitoring of network state and system Zhang, “Interpretable learning algorithm based on XGBoost for fault
prediction in optical network,” in Optical Fiber Communication
performance, and flexible allocation of network resources, in the future. Conference (Optical Society of America, 2020), paper Th1F. 3.
[14] J. Li, D. Wang, M. Zhang, and S. Cui, “Digital twin-enabled self-evolved
ACKNOWLEDGEMENT optical transceiver using deep reinforcement learning,” Opt. Lett., vol. 45,
no. 16, pp. 4654-4657, 2020.
This work was supported by National Natural Science [15] D. Wang, Y. Song, J. Li, J. Qin, T. Yang, M. Zhang, X. Chen, and A. C.
Foundation of China (No.61871415, No.61671076, No. Boucouvalas, “Data-driven optical fiber channel modeling: a deep
learning approach,” IEEE J. Lightwave Technol., vol. 38, no. 17, pp.
61705016) and Fund of State Key Laboratory of IPOC 4730-4743, 2020.
(BUPT) (No. IPOC2020ZT05), P. R. China.