Download as pdf or txt
Download as pdf or txt
You are on page 1of 24

Machine Learning Assisted EDFA Modelling for

Gain and ASE Prediction

Student Name: Kshitiz Singh


Roll Number: 2018154

BTP report submitted in partial fulfillment of the requirements


for the Degree of B.Tech. in Electronics & Communication Engineering
on 6th May 2022

BTP Track: Research

BTP Advisor
Dr. Abhijit Mitra

Indraprastha Institute of Information Technology


New Delhi
Student’s Declaration

I hereby declare that the work presented in the report entitled Multiband Programmable
Optical Amplifier submitted by me for the partial fulfillment of the requirements for the
degree of Bachelor of Technology in Electronics & Communication Engineering at Indraprastha
Institute of Information Technology, Delhi, is an authentic record of my work carried out under
guidance of Dr. Abhijit Mitra. Due acknowledgements have been given in the report to all
material used. This work has not been submitted anywhere else for the reward of any other
degree.

.............................. Delhi, 6th May 2022


Kshitiz Singh

Certificate

This is to certify that the above statement made by the candidate is correct to the best of my
knowledge.

.............................. Delhi, 6th May 2022


Dr. Abhijit Mitra

2
Abstract

Machine Learning methods and algorithms have found use in a vast number of industries, in-
cluding the optical networks industries. One of the key components of any optical network is the
Erbium Doped Fibre Amplifier, or EDFA. EDFAs have variable wavelength dependent output
gain profiles, and are a major source of noise and nonlinear impairments. Thus, EDFA mod-
elling becomes necessary for optical network provisioning and updation, and machine learning
methods are the perfect tool for the same. By means of this study, first, a comprehensive and
highly accurate EDFA model is created by solving the coupled signal propagation equations of
the EDFA. Since the model is complex, and there are vast numbers of channel configurations,
varying numbers of spans as well as EDFA defects, an ML model is later trained to predict the
output gain as well as ASE noise profiles of the EDFA, given the input channel configurations
and the associated pump powers.

Keywords: EDFA, Machine Learning, Optical Networks, ASE Noise, Gain Profile, Runge-Kutta
Method
Acknowledgments

I would like to express my sincere gratitude to my supervisor, Dr. Abhijit Mitra for providing
invaluable guidance, comments and suggestions throughout the course of the project. I would
also like to thank Mr. Rana Kumar Jana for being in touch with me constantly and resolving
any doubts I may have, in addition to providing excellent resources.

i
Contents

1 Introduction 1

2 Basics of EDFA 3
2.1 Basic Physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 3–Level Rate Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 ASE Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 The Signal Propagation Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 The analytical EDFA Model 7


3.1 The Euler Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 The Runge-Kutta Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.3 Simulations and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3.1 Gain v Wavelength . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3.2 Effect of Pump Power on Gain . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3.3 Effect of Channel Loading on Gain . . . . . . . . . . . . . . . . . . . . . . 11
3.3.4 Effect of Number of EDFAs on ASE Noise . . . . . . . . . . . . . . . . . . 12
3.3.5 A Note on Model Complexities . . . . . . . . . . . . . . . . . . . . . . . . 13

4 The ML based EDFA Model 14


4.1 Gain Prediction Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2 ASE Prediction Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.3 Simulations and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

5 Conclusion and Future Work 17

ii
Chapter 1

Introduction

The advent in technology and the onset of Industry 4.0 has brought about new expectations and
higher requirements, especially in the optical network domain. With more and more date being
transferred around the globe each day, the communication systems are getting more complex and
in need of higher maintenance as well. One of the principle components in any communication
system is the amplifier. It is a key device, made necessary as all signals lose energy in the various
media they travel, thus needing amplification so they can be effectively received and understood.
In optical communication networks, the Erbium Doped Fiber Amplifier (or EDFA) is the most
commonly used device for that purpose. The EDFA is so useful as it effectively works in the
C-Band communication range (1530–1565 nm) and offers very low losses.
However, like any physical device, the EDFA is not perfect. It is a large source of noise, as well
as non-linear impairments in the communication channels. Thus, EDFAs have a huge effect on
the quality of transmission (QoT) of signals or the optical signal to noise ratio (OSNR) at the
receiver. Hence, accurate and effective computational modelling of the EDFA becomes necessary.
A computerized model of the EDFA can be effectively used to calculate the gain it provides to
a single or a set of input channels, as well as the noise it adds to the system. Doing this can
be useful for a multitude of real world tasks like an optical network provisioning between two
places, or updating an existing network. An online model can also be used to calculate or check
EDFA aging effects or EDFA degradations, which might impact the network in the future and
cause disruptions.
The aim of this study is to create a very accurate and robust model of the EDFA, which can be
used to carry out large amounts of simulations depending on the input channel configurations,
the effect of degrading pump powers, or the cases where certain channels get suddenly dropped,
and we can observe how these affect the output gain or the ouput noise of the amplifier. The
model can also be used in a span of EDFAs to replicate real world communication systems.
Finally, as the model becomes complex to solve every time using the numerical methods, a
Machine Learning based model is also introduced. The motivation behind this is to use the
machine learning model in addition to the analytical model, which increases the speed as well
as reduces the task of solving complex mathematical equations every time. Combining the two

1
models also takes the EDFA simulation process completely online, with no requirement of any
physical components or laboratory experiments, which was not observed to be done by anyone
in the literature survey. The paper also introduces a way of solving the signal propagation
equations of the EDFA via the Runge-Kutta method, and creating a python model on the same.

2
Chapter 2

Basics of EDFA

2.1 Basic Physics

The Erbium Doped Fiber Amplifiers (EDFAs) are a class of optical amplifiers that can provide
simultaneous amplification of multiple optical wavelength channels. These are extensively used
in modern day optical communication systems, specially for operation in the C–band (1530–1565
nm) and L–band (1565–1605 nm).
An EDFA typically consists of a Silica fiber doped with Erbium (3+) ions. It also consists of a
pump laser (operating at either 980 nm or 1480 nm), which is used to excite the Erbium ions
to an excited state.

Figure 2.1: 3 Level System for the Erbium ion

The 3–Level energy diagram for the Er3+ ion has been shown above, with ’1’ representing the
ground state and ’2’ and ’3’ representing the excited states corresponding to 1480 nm and 980
nm respectively. Assuming just a 980 nm pump, the Erbium ions absorb energy and get excited
from the ground state to state 3. The transition rate depends upon the previous ion densities

3
of the two states as well as the pump flux and pump cross–section (Φp and σp respectively).
The ions then spontaneously lose energy and come down to stable state 2, with a transition rate
of Γ32 . Hence, a population inversion gets created between the ground state and state 2. The
amplification process is then carried out by the phenomenon of ’stimulated emission’. The input
signal wave interacts with the excited Er3+ ions, bringing them back down to ground state.
The ions lose energy by emitting optical energy at the same wavelength and phase as the input
signal. Since the fiber amplifier is long, the stimulated emission output signal at the beginning
of the fiber travels along its length, thus gaining power while remaining of the same wavelength
and phase, in a cascading manner. The rate of transitions from the state 2 to the ground state
depend on the populations of the levels as well as the signal flux and signal cross–section (Φs
and σs respectively). There is also natural decay between the states, the rate of which is given
by Γ21 .

2.2 3–Level Rate Equations

Considering the above 3–Level Energy system for the Erbium ion, we assume the population
densities for the states 3, 2 and 1 to be N3 , N2 and N1 respectively. As these populations change
in the amplification process, the rate equations for the same are given as follows:

dN3
= −Γ32 N3 + (N1 − N3 )Φp σp
dt

dN2
= −Γ21 N2 + Γ32 N3 − (N2 − N1 )Φs σs
dt
dN1
= Γ21 N2 − (N1 − N3 )Φp σp + (N2 − N1 )Φs σs
dt
In steady–state, all these equations turn out to be 0, as there is no change in population densities.
So,
dN3 dN2 dN1
= = =0
dt dt dt
and the total population (N) is
N1 + N2 + N3 = N

The excited state 3 is often very unstable, and ions rapidly decay to the state 2, meaning the
value of Γ32 is very large. Hence, the above system can be approximated to be a 2–level system,
with almost all the ion density shared between state 1 and state 2.

2.3 ASE Noise

Apart from the stimulated emission due to the presence of the input signal, Er3+ ions also decay
from the 2nd state to the ground state spontaneously. The energy lost by the ions is released
in a radiative form, which does not match the input signal in wavelength or phase. Thus, this

4
spontaneous radiation acts as noise in the amplifier. This noise can also get amplified as the
new ”noise signal” travels along the fiber, in the ending constituting the Amplified Spontaneous
Emission, or ASE Noise. The ASE Noise is the major source of signal degradation in the EDFA,
which impacts the quality of transmission or the OSNR of the communication signal. There are
also non–linear effects that lead to noise being added to the EDFA, but this non–linear noise is
smaller compared to the ASE Noise.

2.4 The Signal Propagation Equations

As mentioned in section 2.2, Erbium ions tend to decay rapidly when excited to the State 3 by
a 980nm pump, and come down to the metastable state 2. Hence, the value of N3 is negligibly
small, which means the system can be reduced to a two level model. Assuming that the Erbium
ions are distributed uniformly in the fiber from radius r = 0 to r = R, the ion density of the
second state (N2 ) can be calculated by solving the rate equations in section 2.2, and be written
as a function of the field powers as:

τ σsa P τ σvaj τ σpa


Ahνs Γs Ps + j Ahνj Γvj Pa (vj ) + Ahνp Γp Pp
N2 = P τ (σvaj +σvej )
τ (σsa +σse ) τ (σpa +σpe )
Ahνs Γs Ps + j Ahνj Γvj P a (v j ) + Ahνp Γp Pp +1

where,
Γi : Overlap factors between light–field modes and Er distribution
A = πR2 : effective area of Erbium distribution
h : Planck’s constant
νi : Frequency of signal ’i’
σia : Absorption cross section at frequency ’i’, and
σie : Emission cross section at frequency ’i’.
The absorption and emission cross–sections for an Erbium doped fiber often vary on the host
glass and host environment. Measurement of the absorption cross–section can be done by
normalized absorption measurements on a homogeneous glass sample, and the emission cross–
section can then be computed using the McCumber Relation:

σ (e) (ν) = σ (a) (ν)e(ϵ−hν)/kT

where ϵ is represents energy of transition from 3d state to 2nd state. The cross–sections thus
obtained are shown in figure 2.2.
Finally, using the expression of N2 obtained, the field propagation equations for the pump (Pp ),
the input signal (Ps ) and the ASE Noise (Pa ) can be obtained in diffrential form as:

5
Figure 2.2: Absorption (solid) and Emission (dashed) cross sections of Er(3+) ion for Al–Ge–Er doped
Silica fiber

dPp
= (N2 σp(e) − N1 σp(a) )Γp Pp − αp(a0) Pp
dz
dPs
= (N2 σs(e) − N1 σs(a) )Γs Ps − αs(a0) Ps
dz
dPA (νj )
= ±(N2 σν(e)
j
− N1 σν(a)
j
)Γs PA (νj ) ± N2 σν(e)
j
Γs hνj ∆νj ∓ αν(a0)
j
Pa (νj )
dz
Each ASE power PA (νj ) comprises a forward and backward travelling component, moving in
(a0)
the +z and –z directions respectively. The parameters αi are the possible background losses
faced by the signals as they travel along the length of the fiber. The set of coupled differential
equations thus obtained can be solved to obtain the values of Pump Power, Signal Power and
ASE Power for each channel at the end of the fiber. The gain for an EDFA is the ratio of the
input signal power and output signal power of the amplifier, which can be computed using the
above equations as well. Meanwhile, the ASE noise powers obtained can be used to calculate
the quality of transmission (QoT) or the Optical Signal to Noise Ratio (OSNR).

6
Chapter 3

The analytical EDFA Model

The differential equations for the propagation of the various signals within the fiber amplifier can
be solved to obtain the necessary gain and ASE information at the end of the fiber. However,
the differential equations are all coupled together, via the formula for N2 , making solving them
not a trivial task. Thus, a couple of numerical methods were applied to solve coupled equations,
and their accuracies and complexities were compared.

3.1 The Euler Method

The Euler Method is a nice approximation technique used for solving a set of coupled differential
equations. It can be considered as a first–order Runge–Kutta method.
Suppose we have a differential equation of the form:

dy
= f (x, y)
dx

The Euler Method makes an assumption that for a very small step size ’h’, the value of the
right–hand side of the equation, f(x,y), remains constant. That is f (x, y) = f (x0 , y0 ) in the
range x0 ≤ x ≤ x1 , where h = x1 − x0 .
With this assumption in mind, an approximate solution for the differential equation at a point
’n+1’ can be calculated as:
yn+1 = yn + hf (xn , yn )

The smaller the value of the step–size ’h’, the lower will be the error in approximation.
Applying the method to our set of coupled differential equations, a small step size ’dz’ is consid-
ered first (say, dz = z1 − z0 ). The right side of our equations can be treated as a function f(P,z).
Now, the algorithm is designed in such a way that the value f (Pz−1 , z − 1)dz is computed first,
for all of the equations together, and these get added to the respective power values at the same

7
step (Pz−1 ) to get the power values for the next step. That is,

Pz = Pz−1 + f (Pz−1 , z − 1)dz

This algorithm is applied iteratively for a small step–size (dz) across the length of the entire
fiber, and is applied for all of the equations together. The initial values for the signals (P0 ) at
the beginning of the fiber are sent to the algorithms as parameters. These include the input
Pump–Power, the input Signal–Power and the input ASE–noise Power.

3.2 The Runge-Kutta Method

The Runge-Kutta methods are also iterative algorithms for solving initial value problems (IVPs)
involving differential equations. The most commonly used RK method is the 4th order Runge-
Kutta method, also known as the ’RK4 method’. Suppose we have our differential equation of
the form:
dy
dx = f (x, y) with the initial value y(x0 ) = y0
Here also, we pick an arbitrary small step–size ’h’, and we can define our solution as:

1
yn+1 = yn + h(k1 + 2k2 + 2k3 + k4 )
6

xn+1 = xn + h

with
k1 = f (xn , yn )
h k1
k2 = f (xn + , yn + h )
2 2
h k2
k3 = f (xn + , yn + h )
2 2
k4 = f (xn + h, yn + hk3 )

For applying RK4 to our set of equations, we firstly treat the RHS of the equations as a function
of z, and all the power variables. That is, we consider our differential equations of the form
dP
dz = f (z, Pp , Ps , PA ). Treating the equations in this way makes applying RK4 easier. The
algorithm first computes the k1 values for all equations, using the initial power values input to
the system (which get sent as parameters to the function). Once the k1 ’s get computed, the k2 ’s
can get computed as:

h kp ks kA
k2i = f (zn + , Ppn + h 1 , Psn + h 1 , PAn + h 1 )
2 2 2 2

Similarly, all other k’s can also be computed, and the same process can be repeated across the
length of the fiber for a small step size h (or, dz).

8
3.3 Simulations and Results

The simulations for both the numerical methods discussed above were carried out using the
python programming language. The parameters used in both algorithms have been mentioned
in the table provided below. These values have been taken from the standard literature available
on EDFA characteristics.

Parameter Value
Planck’s Constant (h) 6.626068x10−34
Speed of Light (c) 3x108
Pump Overlap (Γp ) 0.64
Signal Overlap (Γs ) 0.4
Area of Erbium distribution (A) 1.96x10−11
Total Erbium Concentration (N) 1x1025
Erbium Metastable Lifetime (τ ) 10.8 ms
Pump Wavelength (λp ) 980 nm
Pump Absorption Cross Section (σpa ) 2.7x10−25
Pump Emission Cross Section (σpe ) 0

The values of the absorption and emission cross–sections for the other wavelengths (which are
used for the ASE noise and Signal channels) were derived from the graph in figure 2.2. The
other parameters in the equations of the previous chapter, such as the input Pump Power, Signal
Power and ASE Noise Power were provided as per necessary for the simulations. The graphs
obtained from the simulations, and the simulation results follow.

3.3.1 Gain v Wavelength

The Gain for an Erbium Doped Fiber Amplifier is not flat throughout the C+L band optical
communication spectrum. This is due to to the differences in the absorption and emission cross
(observed in figure 2.2), as the gain is directly dependent on their ratio. The gain can be tuned
to be flat in the C-Band range using special filters, or tuning techniques such as Automatic Gain
Control (AGC).
The simulation results obtained via the Euler and RK4 methods are displayed in figure 3.1. It
can be mathematically computed that the error in approximation for the RK4 method is much
lower than the error for the basic Euler method, and a similar sort of observation is made by the
simulations too. The spectral profile obtained by the RK4 method is closer to actual spectral
measurements for an EDFA, and the Euler method is slightly inaccurate, specially for lower
wavelengths. The time complexity of the Euler algorithm, however is almost 4 times lower for
a single amplifier simulation.

9
Figure 3.1: The spectral gain profile of the EDFAs for a single pump power obtained by the Euler Method
(left) and the RK4 Method (right)

3.3.2 Effect of Pump Power on Gain

The biggest influence on the amount of gain an EDFA can provide is the input Pump Power to
the amplifier. A higher input pump power naturally excites a larger amount of Er3+ ions to
the excited state, thus creating a better population inversion, which in turn means the signal
radiation obtained via stimulated emission also has a larger power. The same effect was shown
by simulating both our numerical models, showing they are accurate in physical modelling.
Figure 3.2 shows the spectral profiles obtained for a set of 5 different pump powers. It can also
be observed that both models start to fail at low pump powers (in the range of 4mW), since such
a low power is not enough to create a suitable population inversion between the two Erbium
energy levels.
Figure 3.3 shows the effect of changing the pump powers on the gain for two signals, at 1530
nm and 1550 nm. Here, two observations are the most important. Firstly, the gain severely
drops at lower pump powers, as pointed out above. Secondly, the gain starts to saturate at very
high levels of pump power. This gain saturation effects starts to happen once the pump power
becomes sufficiently high so as to cause a complete population inversion between the two Er
energy levels. At this point, increasing pump power has no effect on the gain, since already all
the population distribution for the Erbium ion lies in the excited state.

10
Figure 3.2: The spectral gain profile of the EDFAs for multiple pump power obtained by the Euler
Method (left) and the RK4 Method (right)

Figure 3.3: Effect of changing the pump power on the gain of a 1530nm and 1550 nm input signal pumped
at 980 nm

3.3.3 Effect of Channel Loading on Gain

The gain for any fiber amplifier often also depends on the multiplicity of the ions within it. It is
also dependent on the homogeneous or inhomogeneous broadening effects. In EDFAs, the rapid
transport of energy across the different Stark–broadened lines as well as the presence of phonons
(heat exchange) is responsible for homogeneous broadening. In the presence of homogeneous
broadening, a strong input signal extracts all the energy stored in the amplifier. But, this has
a downside as a homogeneously broadened amplifier used in a wavelength division multiplexed
(WDM) system is very susceptible to adding and dropping of wavelength channels which affects
the gain and energy extraction of other channels. This effect was also modelled, and some gain
profiles with a different number of channels have been plotted in figure 3.4.
As it can be observed, adding of more channels to the amplifier infact reduces its overall gain
across channels. This can be intuitively thought of as that the energy stored within the amplifier
gets distributed across a larger number of channels when more channels gets added, leading to

11
lower gains observed for certain channels as compared to when a lesser number of channels are
on. The key takeaway from figure 3.4 should be the comparative values of the gains observed
at similar wavelengths for the different channel loading cases, which highlight the point made
above.

Figure 3.4: Gain profiles observed for different channel loading configurations keeping other parameters
fixed

3.3.4 Effect of Number of EDFAs on ASE Noise

As we saw in the previous chapter, EDFAs add significant amount of linear noise to the com-
munication system, in the form of the Amplified Stimulated Emission Noise. Naturally, as the
number of EDFAs in a span or link increase, the noise travelling through the link also increases.
To simulate this effect, multi-span links containing from 1 to 8 EDFAs were simulated. It was
assumed that each EDFA is 80km apart, and the signals travelling through the connecting opti-
cal fibers suffer an attenuation of 0.25 dB/km. Thus, to compensate the fiber losses, each EDFA
was programmed to have a gain of approximately 20dB (by selecting a suitable input pump
power), and the signals were generated and propagated through the entire link. The total noise
obtained at the end of the link was then plotted against the wavelength for the different number
of EDFAs in the link, in figure 3.5.

12
Figure 3.5: ASE Noise per wavelength channel for different number of EDFAs in the link. EDFAs were
all 80km apart, with gain of 20dB

As can be observed, increasing the number of EDFAs progressively increases the amount of noise
going through the system. The spectral profile of the noise powers also follows the gain–profile
like shape, since the noise at the start of the fiber was assumed to be zero, and the only source
of noise in the system was assumed to be EDFA ASE noise.

3.3.5 A Note on Model Complexities

Naturally, the Runge-Kutta algorithm is more time complex than the Euler algorithm, since
the Euler method is basically a first–order RK method. Noting the simulation times for both
the models, the RK4 algorithm was roughly 4 times complex than the Euler one, when just a
single EDFA was being simulated. However, increasing the number of EDFAs to ’n’ made the
RK-4 algorithm taking longer and longer to finish, since its complexity effectively turned out to
be approximately 4n times that of the Euler method. The positive tradeoff for the same being
lower errors and better accuracies in the physical modelling of the EDFA by the RK4 method.

13
Chapter 4

The ML based EDFA Model

Since most practical optical communication systems can contain upto a few hundreds of wave-
length channels with multiple EDFAs in a link, running simulations using the analytical model
(using the RK4 method) becomes less feasible due to its high complexity. To illustrate the
same, a graph was plotted between the time taken to simulate all possible channel loading
configurations for a system with 10 channels, and the number of EDFAs in a link.

Figure 4.1: Time taken to run a set of 1024 10–channel simulations for different number of EDFAs using
the RK4 algorithm.

It can be observed how the time complexity increases exponentially in the figure 4.1. Thus, the
need for a Machine Learning model to make this task faster and more robust arises. An accurate
ML model trained on sufficient data can effectively be used to predict the gain or ASE noise
for any EDFA or link of EDFAs in a much smaller time than running all analytical simulations.
For the same, two neural networks (NNs) were designed: one for gain–prediction trained on
all possible channel configurations for 10 wavelength channels in a system with upto 4 EDFAs,
and another for the ASE Noise prediction in a system with upto 133 channels and 8 EDFAs.
The second model was designed considering practical sprectral–assignment tasks in real–world

14
optical networks. A sample neural network architecture with 2 hidden layers is shown in figure
4.2. Similar architectures were used in both the following models, containing different numbers
of neurons per layer as required by the datasets.

Figure 4.2: Sample Neural Network Architecture with 2 Hidden Layers

4.1 Gain Prediction Model

The first gain prediction model was designed to work on a system with at most 10 wavelength
channels, and at most 4 EDFAs in a multispan link. The model was capable of predicting the
channel–wise gain for a communication system containing any possible ”on–off” configuration
of the channels, as well as the EDFAs being in different degrading configurations. To explain
the degrading configurations, assume there is a communication link with 4 EDFAs in between.
Now, due to ageing or external degradation effects, one or more, or even all of the EDFAs could
start degrading, meaning their output gain starts reducing. These EDFAs could be in different
stages of degradation as well, like say EDFA-1 in the link now gives a gain of just 10 dB while
EDFA-3 gives a gain of 16 dB. So, a very large amount of such cases were generated using the
RK4 model, and the datset hence formed was used to train the NN.
The neural network architecture consisted of an input layer with 16 neurons, two hidden layers
with 15 neurons each, and an output layer with 10 neurons for the gain profiles of the 10 channels.
The input layer comprised of the number of EDFAs in the link, their degradation configurations
(1: if an EDFA is degrading, 0 otherwise), the input pump powers for the EDFAs as well as the
10 channel loading configuration (1: if a channel is ON, 0 otherwise). The dataset generated
had a total of 189,255 data points, which were split in a train–validation split of 0.8:0.2 .

15
4.2 ASE Prediction Model

The second, ASE prediction model was designed to work on a system containing as many as
133 wavelength channels, with anywhere between 1 to 8 EDFAs in the link. Once again, various
degradation states for the EDFAs were generated, by providing each EDFA a random input
pump power. Here, the channel loading configurations were not completely random. Since in a
practical real–world system, spectrum allotment for lightpath requests in not done randomly, but
in a contiguous channel block, the same was modelled while generating the dataset. A random
number of channels ranging from 1 to 133 (all full) were generated, and they were assigned a
random contiguous block within the spectrum. The number of EDFAs wee also randomized, as
well as their input pump powers.
The neural network architecture for this model consisted of an input layer with 143 neurons,
2 hidden layers with 143 neurons each, and an output layer with 133 neurons (for the ASE
noise for each channel). The input layer comprised of the number of EDFAs, the number of
ON channels, the input pump powers for all EDFAs, and the channel configuration (1: channel
is on, 0 otherwise). The generated dataset had 10,000 data points. Since the simulations for
133 channel cases were very complex to run, a lesser number of data points were generated.
However, since there was no need for the randomization of channel configurations this time, a
very large dataset was also not really required. Similar as the last dataset, the train–validation
split was in a 0.8:0.2 ratio.

4.3 Simulations and Results

Once the datasets were generated, basic dataset preprocessing techniques like min–max normal-
ization were applied. NaN values were also removed. Since the dataset for the 10 channel gain
prediction model was very large in size, the model training was carried out only for 10 epochs
for it. The number of epochs chosen for the smaller 133 channel ASE Noise dataset were 100.
Suitable low learning rates (of 0.02) were chosen for both the model trainings, and the activation
function for each layer was ”relu”. Stochastic Gradient Descent was chosen as the optimizer,
and the loss function was mean–squared error. The training and testing MSE values obtained
for both the datasets are mentioned in the following table.

Model Training Error (dB) Testing Error (dB)


10 Channel Gain Prediction 0.23639 0.23460
133 Channel ASE Prediction 0.34898 0.34732

The errors obtained for the ML prediction tasks for both the models came out significantly low
(less than 0.5 dB), meaning the Machine Learning models can estimate the gain or ASE Noise
from a link of EDFAs with multiple channels in a very effective manners. Once the weights of
the model are stored, there remains no need to train the model further, and the dependency on
the complex RK4 method is completely removed for any future simulations.

16
Chapter 5

Conclusion and Future Work

By means of this study, I have come up with a machine learning assisted fully computerized
model for running simulations on the Erbium Doped Fiber Amplifier. The model introduced
can run all forms of simulations with respect to the EDFA, based on the number of channels
input, the number of EDFAs in a link, the input pump power etc. to very efficiently predict
the Gain as well as the output ASE noise from the device(s). Many of the effects of these
parameters on the Gain of the EDFA were also studied and plotted, which were in accordance
with the results published in other literature. The initial analytical model was based on solving
the coupled differential equations for signal propagation in EDFAs by using the Runge Kutta
Method. A couple of neural network based ML models were also introduced, that were trained
on the data generated by the analytical model which makes the overall gain and ASE prediction
task effecient, highly effective and robust.
The major importance of a fully online EDFA model like this is that it completely takes any
dependency off physical devices and laboratory experiments. The model can also be extended
to do a variety of tasks, like creating efficient RMSA policies, or solving tasks like soft-failure
detection. Since ASE noise can easily be computed using this model, it can also be used in tasks
based in QoT estimation or OSNR prediction.

17
Bibliography

[1] P.C. Becker, N.A. Olsson, J.R. Simpson, Chapter 5 - Erbium-Doped Fiber Ampli-
fiers—Amplifier Basics, Editor(s): P.C. Becker, N.A. Olsson, J.R. Simpson, In Optics and
Photonics, Erbium-Doped Fiber Amplifiers, Academic Press, 1999, Pages 131-152, ISSN
15575837, ISBN 9780120845903, https://doi.org/10.1016/B978-012084590-3/50007-7.

[2] P.C. Becker, N.A. Olsson, J.R. Simpson, Chapter 6 - Erbium-Doped Fiber Ampli-
fiers—Modeling and Complex Effects, Editor(s): P.C. Becker, N.A. Olsson, J.R. Simpson,
In Optics and Photonics, Erbium-Doped Fiber Amplifiers, Academic Press, 1999, Pages
153-200, ISSN 15575837, ISBN 9780120845903, https://doi.org/10.1016/B978-012084590-
3/50008-9.

[3] Desurvire, Emmanuel and Michael Zervas. “Erbium-Doped Fiber Amplifiers: Principles and
Applications.”

[4] Jiakai Yu, Shengxiang Zhu, Craig L. Gutterman, Gil Zussman, and Daniel C. Kilper,
”Machine-learning-based EDFA gain estimation [Invited],” J. Opt. Commun. Netw. 13, B83-
B91 (2021)

[5] A. Mahajan, K. Christodoulopoulos, R. Martı́nez, S. Spadaro and R. Muñoz, ”Modeling


EDFA Gain Ripple and Filter Penalties With Machine Learning for Accurate QoT Estima-
tion,” in Journal of Lightwave Technology, vol. 38, no. 9, pp. 2616-2629, 1 May1, 2020, doi:
10.1109/JLT.2020.2975081.

[6] S. Shahkarami, F. Musumeci, F. Cugini and M. Tornatore, ”Machine-Learning-Based Soft-


Failure Detection and Identification in Optical Networks,” 2018 Optical Fiber Communica-
tions Conference and Exposition (OFC), 2018, pp. 1-3.

[7] Y. You, Z. Jiang and C. Janz, ”Machine Learning-Based EDFA Gain Model,”
2018 European Conference on Optical Communication (ECOC), 2018, pp. 1-3, doi:
10.1109/ECOC.2018.8535397.

[8] Y. You, Z. Jiang and C. Janz, ”OSNR prediction using machine learning-based EDFA mod-
els,” 45th European Conference on Optical Communication (ECOC 2019), 2019, pp. 1-3,
doi: 10.1049/cp.2019.1044.

18
[9] Christian Merkle. 2010. Degradation model for erbium-doped fiber amplifiers to reduce net-
work downtime. In Proceedings of the 16th EUNICE/IFIP WG 6.6 conference on Networked
services and applications: engineering, control and management (EUNICE’10). Springer-
Verlag, Berlin, Heidelberg, 198–208.

19

You might also like