Optimal Design of Raman Amplifier

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Optimal design of Raman amplifier

J.K. Behera*a, R.K.Shevgaonkara


a
Dept. of Electrical engg, IIT Bombay, Powai, Mumbai, INDIA 400076

ABSTRACT
Raman amplifier can be thought as more of a loss compensator than an amplifier. Hence we normally expect all the
signals to come out of the amplifier with same signal strength with which they entered it. For that we ideally require a
flat gain profile. By adjusting the powers and wavelengths of multiple pumps employed this profile is achieved.
Obtaining this profile is a computationally expensive and difficult task. The reason being the highly non-linear and
multi-modal nature of the system involved. Inside the distributed Raman amplifier, which is nothing but an arrangement
of long transmission fiber and few high -power light sources, signals are affected by a number of factors, like, fiber
losses, Rayleigh back-scattering (RBS), double RBS, amplified spontaneous emission (ASE) noise, and most
importantly, stimulated Raman scattering. When all these factors considered together, one would find it almost
impossible to express the underlying mathematical nature of the system analytically thereby failing to analyze it's
behavior. In this thesis, an efficient method to achieve a flat-gain profile has been proposed. Usage of genetic algorithm
facilitates the optimization problem where the objective is to obtain a suitable combination of wavelength and power for
each source, which would result in minimal gain ripple. Issues like pump power space quantization and LASER power
effect has been discussed.
Keywords: Optimization, genetic algorithm, Raman amplifier

1. INTRODUCTION
1.1 Overview of Raman amplifiers
The advent of the optical amplifier changes the optical fiber communication systems and invites rapid development of
wavelength-division-multiplexed (WDM) transmission systems and networks. Although the entire band of erbium -
doped fiber amplifier (EDFA) is fully utilized, the need for more optical channels along installed fibers, increasing
channel count and optical bandwidth, p ushes EDFA technology beyond its performance limit.
To increase WDM data capacities and realize the ultra-long-haul system, the noise performance of amplifiers becomes
critical. Although the capacities can be upgraded by means of reducing the WDM channel spacing, the nonlinear effects
in the transmission fiber such as cross-phase modulation (XPM) and four-wave mixing (FWM) starts to impair the
optical waveform significantly. And although these nonlinear effects can be suppressed by reducing the optical signal
power launched into the fiber, in turn, the transmission distance is limited due to optical signal-noise-ratio (OSNR)
degradation. Fortunately, distributed Raman amplifier (DRA) makes the noise performance significantly improve,
longer transmission dis tances be realized and existing systems be upgraded. What is more, DRA can mitigate fiber
nonlinear effects, and improve the OSNR without the need to increase the input optical signal power. Two critical
merits of DRA are the low noise and the arbitrary gain band.
1.2 Raman amplification
Raman gain arises from the transfer of power from one optical beam to another that is downshifted in frequency by the
energy of an optical phonon. Raman gain spectrum in fused silica fibers is illustrated in Fig. 1. The gain bandwidth is
over 40 THz wide, with the dominant peak near 13.2 THz. The gain band shifts with the pump spectrum, and the peak
value of the gain coefficient is inversely proportional to the pump wavelength. In the telecommunications bands are
around 1500nm.

Raman amplifiers have some fundamental advantages. First, Raman gain exists in every fiber, which provides a cost-
effective means of upgrading from the terminal ends. Second, the gain is non-resonant, which means that gain is
available over the entire transparency region of the fiber ranging from approximately 0.3 to 2µm. A third advantage of
Raman amplifiers is that the gain spectrum can be tailored by adjusting the pump wavelengths. For instance, multiple
pump lines can be used to increase the optical bandwidth, and the pump distribution determines the gain flatness.
However, a number of challenges for Raman amplifiers prevented their earlier adoption. First, compared to the EDFAs,
Raman amplifiers have relatively poor pumping efficiency at lower signal powers. Although a disadvantage, this lack of
pump efficiency also makes gain clamping easier in Raman amplifiers. Second, Raman amplifiers require a longer gain
fiber. However, combining gain and the dispersion compensation in a single fiber can mitigate this disadvantage. A
third disadvantage of Raman amplifiers is a fast response time, which gives rise to few sources of noise. Finally, there
are concerns of nonlinear penalty in the amplifier for the WDM signal channels.

Fig. 1: Raman gain spectrum for SMF Fig.2: Schematic of Raman amplifier

1.3 Distributed Raman amplifier


Figure 2 shows schematic of a fiber Raman amplifier. Fiber Raman amplifiers typically consist of an optical fiber as the
gain medium, a counter-propagating pump light source, a wavelength-sensitive coupler to combine the pump and the
signal, and an optical isolator. With the exception of the gain medium, the configuration resembles that of an EDFA. In
Raman amplification, it is important to use un-polarized pump sources to avoid polarization-dependent gain, because
stimulated Raman scattering is a polarization-sensitive process. Depolarization, in particular, is necessary for laser
diodes that emit linearly polarized light.

Counter-pumped, distributed Raman amplifiers can achieve better signal-level excursion because the signals tend to be
amplified only when their power level is lowered because of fiber span loss. In Fig. 3, in which a counter-pumped
Raman gain is assumed, it can be seen that the average path signal level is low, allowing nonlinear effects to be avoided,
although the signal level fits well in the admissible transmission window. To resolve various noise issues, Raman
amplifiers are usually configured as counter-pumped. The use of co-pumping along with counter-pumping has other
advantages. First, it helps to reduce the total launched pump power from the counter-pumping light source. This is
particularly important for safety. Second, co-pumping, compared with counter-pumping, yields further enhancements,
with relatively low pumping power, in the OSNR. It should be noted, however, that too much co-pumping also increases
non-linearity. Third, balanced, bi-directional pumping best suppresses DRBS-MPI noise.

Fig. 3: Signal strength variation comparison of DRA & EDFA

Figure 3 plots signal power variations over distance with and without a distributed gain, showing how distributed gain
enhances transmission performance. The curve shows variation of signal power in case of Raman amplifier counter-
pumping and the down-sloping line is the case of conventional pumping. In designing an optical transmission system,
one should optimize the signal level in terms of noise and non-linearity; optical amplification of lower signal levels adds
higher noise, whereas higher-level signals suffer more from non-linearity in fiber. It is apparent from the figure that
introducing a distributed gain increases the range of the optimal signal levels.

In a WDM pumping scheme, one combines a plurality of pump lasers, with different center wavelengths, to launch into
the optical fiber at the same time. The WDM pump light waves interact with each other through stimulated Raman
scattering, transferring the energy from short to long wavelengths. For this reason, the path-average, or effective, pump
power at each wavelength is different from the launched power, to an extent, which depends on the strength of the
interaction. However, because each of the stimulated Raman scatterings caused by different pump wavelengths occurs
independently, the composite gain that is created is simply the sum of every logarithmic gain proportional to the
effective pump power at each wavelength. The number and allocations of the wavelengths depend on the gain band and
on the degree of flatness required in the system. The pump wavelength range determines the gain band.

A wide, flat composite gain can be constructed by creating a pair of opposite gain slopes: because the spectral shape of
Raman gain has a slope that is less steep around the peak on the shorter wavelength side, it is necessary to prepare many
small gains, slightly shifted from each other, to form jointly a smooth spectral slope opposite, around the peak on the
longer wavelength side.

2. OPTIMAL DESIGN PROBLEM

Power variation of signals and pumps in the backward pumped fiber Raman amplifier with multiple signal and pump
channels can be described by the following system of coupled nonlinear equations.
n +m
dPk
±
dz
= −α k Pk + ∑ g (υ
j ≠k , k =1
j ,υ k )Pj Pk (2.1)

for k=1,2,…,n+m
where n is the number of pump waves, m is the number of signal channels, and values Pk , υ k and α k describe,
respectively, the power, the frequency and attenuation coefficient for the k th wave. Frequencies are numbered in
decreasing order (υi > υ j , i<j). The plus (minus) sign on the left of equation 2.1 corresponds to forward k=n+1,…n+m
(backward k=1,2,….n) propagating waves. For fiber span of length L the boundary conditions are defined at z=L for
pumps and at z=0 for signals. The conditions are, Pk (L)= Pk 0 (k=1,2,….n) and Pk (0)= Pk 0 (k=n+1,…n+m). The gain
coefficients g (υ i ,υ j ) describe the power transfer by stimulated Raman scattering (in Stokes direction) between jthand
kth waves given by:
g (υ j ,υ k ) =(1/K effAeff)gj(υ j − υ k )
for υ j > υ k and
g (υ j ,υ k ) = -(1/KeffAeff) (υ j / υ k ) gj( υ k − υ j )
for υ j < υk ,
where gi(∆ν) is the Raman gain spectrum measured at the pump frequency ν i, Aeff is the effective core area of the fiber
and K eff is the polarization factor. The aim is to solve the equation (2.1) by taking Raman gain spectrum and the
attenuation spectrum of the fiber as the given a priori characteristics in order to determine the optimal pump powers and
pump frequencies. Also, one should specify the amplifier el ngth, the input signal level, the wavelength range for signal
and for pump waves, and the number of signal and pump channels m and n, respectively .

Now equation 2.1 can only be solved numerically by any iterative method. We can see that, for n pumps we have n
pump powers and n frequencies, i.e., 2n variables to adjust in order to achieve a flat -gain spectrum. Here the
optimization problem is to get the suitable combination of pump powers and frequencies, which would result in equal
amplification of all signal channels with number of signal and pump channels given. As the mathematical model, or the
relationship between output signal strength to input signal strength can’t be represented analytically, it makes the usage
of any simple optimization algorithm almost impossible or extremely time-consuming. As the underlying mathematical
function is not differentiable, application of gradient methods to this optimization problem is impossible. So we propose
to use genetic algorithm, which finds solution to this problem faster and more accurate than any other algorithm.
Advantage of this over other algorithms is it does not need to know what the mathematical nature of the system is. The
function can be multi-modal, non-linear, non-differentiable etc. In each case genetic algorithm works elegantly and
proves its robustness.
3. GENETIC ALGORITHMS: AN OVERVIEW

A genetic algorithm is a heuristic search algorithm for the solution of optimization problems in which, starting from a
random initial guess solution, better descendants are tried in an attempt to find one that is the best under some criteria
and conditions. It is based on the idea of evolution theory that individuals having a high value of quality will survive to
the next generation with greater probability.

Algorithms for function optimization are generally limited to convex regular functions. However, many functions are
multi-modal, discontinuous, and non-differentiable. Genetic algorithms have been used to solve difficult problems with
objective functions that do not possess nice properties such as continuity, differentiability, satisfaction of the Lipschitz
Condition, etc.[5]. These algorithms maintain and manipulate a family, or population, of solutions and implement a
survival of the fittest strategy in their search for better solutions. Genetic algorithms are very efficient when the
solution space is very large and the function is highly non-linear.

Genetic algorithms search the solution space of a function through the use of simulated evolution, i.e., the survival of
the fittest strategy. In general, the fittest individuals of any population tend to reproduce and survive to the next
generation, thus improving successive generations. However, inferior individuals can, by chance, survive and also
reproduce. Genetic algorithms have been shown to solve linear and nonlinear problems by exploring all regions of the
state space and exponentially exploiting promising areas through mutation, crossover, and selection operations applied
to individuals in the population. A more complete discussion of genetic algorithms, including extensions and related
topics, can be found in the book, Goldberg[5]. A simple genetic algorithm can be described as below.

1. Generate a population P0 on N random individuals and respective values of function for each of the individuals.
2. i=1
3. Pi’= selection(Pi-1)
4.Pi’=crossover(P’i)
5. Pi’=mutation(Pi’)
6. i=i+1
7. Go to 3, until termination criterion is met.
The use of a genetic algorithm requires the determination of six fundamental issues: chromosome representation,
selection function, the genetic operators making up the reproduction function, the creation of the initial population,
termination criteria, and the evaluation function.

3.1 Chromosome representation


Chromosome or an individual is a collection of input variables whose optimal values will minimize or maximize the
function to be optimized. Suppose, f(x1,…xn) is to be optimized then an individual or a chromosome means a collection
of x1,….xn with each value initialized to some random value between range of each xi. This is the case of real-coded
genetic algorithm, in case of binary GA, one can represent the individuals as binary equivalent of each variable. One can
also go for integer-coded GA.
Fig 4. Generating initial pop ulation for 0≤xj ≤10, j=1,…,5

For any GA, a chromosome representation (Fig 4) is needed to describe each individual in the population of interest.
The representation scheme determines how the problem is structured in the GA and also determines the type of genetic
operators that are used. Originally only binary-coded GA was used. It has been shown that more natural representations
are more efficient and produce better solutions. One useful representation of an individual or chromosome for function
optimization involves genes or variables from an alphabet of floating-point numbers with values within the variables
upper and lower bounds. This is called real-coded GA. Advantage of a real-coded GA is that; it's an order of magnitude
more efficient in terms of CPU time. It's been found [6] that a real-valued representation moves the problem closer to
the problem representation, which offers higher precision with more consistent results across replications.

3.2 Selection function


Selection function is the driving force behind genetic algorithms. The selection of individuals to produce successive
generations plays an extremely important role in a genetic algorithm. A probabilistic selection is performed based upon
the individual's fitness such that the better individuals have an increased chance of being selected. An individual in the
population can be selected more than once with all individuals in the population having a chance of being selected to
reproduce into the next generation. There are several schemes for the selection process: roulette wheel selection and its
extensions, scaling techniques, tournament, ranking methods [5].
A common selection approach assigns a probability of selection, Pj, to each individual, j based on it's fitness value. This
fitness value is determined using a fitness function, which is in general the function, which is to be maximized or
minimized itself. A series of N random numbers is generated and compared against it cumulative probability, C i
i
= ∑ Pj of the population. The appropriate individual, i, is selected and copied into new population if Ci-1<U(0,1)<C i.
j =1
Here U(0,1) is a random number between 0 and 1. Few popular methods to assign probabilities are roulette-wheel
selection, ranking-methods, and tournament selection.
3.3 Crossover
Genetic operators like crossover and mutation constitute the basic search mechanism of genetic algorithms. These
operators create new solutions based on existing ones. Crossover takes two individuals (parents) and produces two new
individuals (offspring). There are many types of crossover. Type of crossover to be used is different for different type of
representation of individuals. If X and Y are two m-dimensional row vectors then simplest crossover is performed by
generating a uniform random number r between 1 and m and creating offspring X’, Y’ with elements of X’ made equal
with the corresponding elements from Y’ if, i>r and vice-versa. Elements are unchanged if i<r.
3.3 Mutation
Crossover produces better individuals from the existing population whereas mutation performs a local search for
optimal solution by picking a variable randomly and changing its value to find a new solution, which might be optimal.
Figure shows the mutation process.
Fig. 5: Example of crossover Fig. 6: Example of mutation

3.4 Strategy for optimization


The underlying theory behind genetic algorithms resembles Darwin's principle of survival of the fittest, meaning, out of
the large population of random samples, those who are fit, will move from generation to generation and those who don't
fit into the fitness criterion will die down. The only thing one needs to do is to find a suitable definition of fitness. The
fitness function is most of the cases the function itself which is to be optimized. The fitness term can be the output of the
function itself (in case of maximization) or a scaled version of it (in case of minimization). Though most of the cases
GAs require ordered set of non-negative fitness values (because we calculate probability out of it, which later is
compared against a positive random number), that does not limit the scope of GAs in minimizing or maximizing of
functions which take negative values. One simple example of fitness scaling will be like this;
Fitness (scaled) =C max - f( x )

Here Cmax is a constant which would be much greater than possible maximum of the function to be optimized, f is the
function which is to be optimized, x is a random sample solution. So, when function approaches it's minimum, scaled
fitness will approach maximum and vice versa. Hence we can also minimize functions in addition to maximizing by
fitness scaling. It is same for the case for functions, which take both negative and positive values.
We apply three evolution procedures, visually, selection, crossover, and mutation repeatedly in each iteration.
Selection process finds out fitness of each individual, and picks up copies of fittest individuals randomly, crossover
performs mating of randomly chosen pairs to produce new (usually better) individuals, and mutation does a local search
by making small random changes in offspring. Eventually, GA forces all individuals of the population to converge to a
single solution.
The GA must be provided with a random population, which is valid, valid in the sense, that, all members of each
individual must lie within the range of the actual solution. Point to be noted here is, as initial population is a collection
of random numbers, each time we run the ge netic algorithm, it's possible to end up with different solution each time.
However, it's been observed that all these solutions come out to be very close and highly competing in nature. It's
always better to seed initial population with few better individuals and rest of the population with randomly chosen
ones. It will facilitate faster convergence.

4. SIMULATION PARAMETERS & METHODOLOGY

Power in each signal channel was kept at 0dBm. Amplifier length was kept to be L=50km, with K eff=2, Aeff, which is
effective core radius area of silica fiber, was assumed to be 50x10-12mm2, the losses for signals channels were assumed
to be 0.2dB/km and for pumps it was assumed to be 0.35dB/km. Using shooting algorithm [2], for each combination of
pump power and frequencies, signal gain spectrum was found out. Here the pumping power of each pump was allowed
to be any value between 0 to 300mW. The system was simulated for 10, 20, 30 channels with 1, 2, 3 pumps. The genetic
algorithm has been implemented as follows.
• Integer-coded GA has been used with each chromosome containing integer pump powers and wavelengths.
Wavelengths are converted into frequencies and fitness of each chromosome is evaluated.
• Selection function uses tournament selection.
• Because of the integral nature of the chromosomes, discrete recombination as crossover.
• Range of pump powers was quantized, i.e. available pump powers were given as integers in steps of 1, 10 and
20 and the corresponding gain ripple was calculated and comparison of these results are given in the following
section. This is needed as LASERs with high power accuracy is a difficult task and manufacturing pumps with
integral powers would be more convenient.
LASER power drift has been taken into account in order to calculate the perturbation in gain-flatness with drift of the
pump power sources.

5. RESULTS

The gain ripple used here is the difference between maximum and minimum deviation when some of the signals have
gain below 0dB line and some have above 0dB line. If all of the signals have gains below 0dB line, then gain ripple is
absolute of the minimum of all these gains. If all the signals have gain above 0dB gain line, then gain ripple is equal to
maximum of all. For 1,2,3 pumps and for 10,20,30 channels, the following table shows the gain ripple.

No. of channels No of pumps Gain-ripple (dB)


10 1 0.0686
3 0.1224
20 1 0.269
3 0.2386
1 0.5121
30 2 0.4904
3 0.3615
Table 5.1: Variation of gain ripple with number of pumps used

Quantization step of pump power (mW) Gain ripple (dB) No. of generations taken by GA
1 0.2410 33
10 0.3458 21
20 0.4378 20
Table 5.2: Pump power space quantization

With time power of the LASER sources fluctuate. It has been termed as power drift. To design a practical Raman
amplifier we need to take ht is factor into account to analyze the sensitivity of the gain ripple in the amplifier with the
power drift. This is because of the following fact. Suppose we run the simulation process to get a set of pump frequency
and power combinations. Now as the power of each of the pumps change with time, and if, with this variation of the
powers, the gain ripple of the amplifier changes drastically, then it's hardly useful. So varying the pump powers from
their optimized values by ±5%, ±2%, ±1% and applying genetic algorithm to analyze this, we get the result as shown in
table 5.4. Here, a case of 3 pumps and 30 signal channel case has been considered. By running the proposed algorithm,
the optimal pump powers thus obtained are shown below.
Pumps used Wavelength (nm) Power (mW)
1 1424 111
2 1465 94
3 1498 124
Table 5.3: Optimized pump power and wavelength combinations

Here the gain ripple comes out to be 0.4318dB, for this particular combinations of pump powers and frequencies. Now
the pump powers were allowed to vary within following ranges (table 5.4) and corresponding worst-case gain ripple was
calculated as follows.
Power drift (±%) Gain ripple (dB)
5 0.8204
2 0.5254
1 0.4388
Table 5.4: Variation of gain ripple with power drift
6. CONCLUSIONS

A brief theory of genetic algorithm and its application to highly non-linear problem of Raman amplifier design was
discussed. The algorithm has been developed which has an advantage over other numerical optimization techniques in
the terms of its time complexity.
As shown in table 5.1, for a fixed number of channels as we increase the number of pumps, the ripple, that is the
difference between maximum and minimum gain increases. Then the question, which can be asked here, is why would
one go for more number of pumps, if the gain-ripple increases as the number of pumps increases! But fortunately, this is
observed in case of lesser number of pump channels, say for example 10 signal channels. The gain ripple increases from
0.0686dB to 0.1224dB, as we vary the number of pumps from 1 to 3. One explanation to this observation could be this.
For a certain number of channels, we can get a certain achievable lower limit of gain ripple. It's not that, we can increase
the number of pumps used arbitrarily to get any value of gain ripple we want. This is because as the number of channels
is less, and the number of pumps increase, the pump-pump and pump-signal Raman interaction becomes very strong and
each signal experiences high gain and all these gains are deviated highly from each other, meaning, the minimum gain,
the maximum gain and all the intermediate gains differ from each other by large values. But when the number of signal
channels increase, we see that this situation is not there. Because, as depicted in a 20-signal channel case, the gain ripple
increases first, then decreases. But, finally, in case of 30 signal channels, we get the ripple decreasing from 0.5121dB to
0.3615dB, because this is a case of poor pumping. Of course the gain ripple would stop reducing further as the number
of pump increases above a certain value.
Then considering the power drift in laser sources, an analysis of the increment in gain ripple was given by varying the
optimized powers by few percent of its ideal value. The analysis in table 5.4 shows, for less than ±2% of power drift the
amplifier works perfectly in the sense that the gain ripple does not increase much and now-a-days laser sources with
power drift less than ±2% are available. Also as laser sources with wavelength stability of ±0.01nm are available, this
factor does not have any severe effect on gain ripple of the amplifier. As the genetic algorithm manipulates random
sample solutions to find out global optimum one, it might not give the best result at one shot. It's always better to run the
algorithm multiple times and to choose the best solution.

REFERENCES

1. G.P.Agrawal, Fiber-optic Communication Systems, Wiley and Sons, Inc., 2002.


2. Effective shooting algorithm and it's application to fiber amplifiers: Xueming Liu and Byoungho Lee, OPTICS EXPRESS, Vol. 11,
No. 12, June 2003.
3. A genetic algorithm for function optimisation, A MATLAB implementation: Christopher R. Houck, Jeffrey A. Joines, Michael G.
Kay, North Carolina State University.
4. On the use of genetic algorithm with elitism in robust and non-parametric multivariate analysis: M. Zirngibl, Austrian Journal of
Statistics, vol.32, No. 1,2, 13-27, April 2003.
5. David E. Goldberg, Genetic algorithms, in search, optimization and Machine learning, International Student Edition, Addison
Wesley Inc. 1999
6. Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs, Artificial Intelligence Series, Springer-Verlag, New
York.

*jiitb@yahoo.com

You might also like