Asia Communications and Photonics Conference 2021 © Optica

T4A.98
Publishing Group 2021

Data-effective and Accurate EDFA Gain Prediction

Black-box Model
Mingming Zhang1, Yuqi Li1, Weihao Li1, Junda Chen1, Yizhao Chen2 and Ming Tang1*
1
Wuhan National Lab for Optoelectronics (WNLO) & National Engineering Laboratory for Next Generation Internet Access System, School of
Optical and Electronic Information, Huazhong University of Science and Technology, Wuhan, 430074, China
2
School of Cyber Science and Engineering, Huazhong University of Science and Technology, Wuhan, 430074, China
*[email protected]

Abstract: An EDFA gain prediction model is proposed. With a small size of experimental data,
the spectral gain root-mean-squared-error of the model is 0.1756 dB.

1. Introduction
Erbium-Doped Fiber Amplifiers (EDFA) have been widely used in optic fiber communication systems for decants
of years. It dominates the range of optic-amplifiers by its wide-band gain feature and high efficiency. With the
increasing demands of information capacity, the bandwidth of spectrum has been extended. When the gain spectrum
becoming wider and the input signal power becoming higher, the gain flatness of EDFA turns to be tilted and
rippled. Especially in the long-haul repeaterless link, the flatness degradation of the spectral gain damages the
transmission performance of WDM channels [1]. It is of great importance to build effective models to help to predict
and compensate the degradation of power distribution.
In order to predict the operating status of EDFA, many simulation models and prediction models were proposed.
Those models can be divided in to two categories by their periods. The older models tend to be analytical and simple
[2,3]. However, their accuracy is limited in wide-band and high-power scenario, because they are not able to make
full use of measurement data. The newer ones are mostly based on the cutting-edge machine learning technology.
And they do realize a real good fit, but their complexity seems quite large [4,5]. Thousands of groups of data will be
needed for a one-shot training. Whenever the operating state of those well-trained EDFA devices altered, another
data-collecting drudgery would be carried out again. The “EDFA training work” can be simplified without cost of
accuracy, when given some priori knowledge of its physical properties. Compared to the data-lavish machine
learning methods, we propose an accurate semi-analytical EDFA gain model, which is simple and effective.

2. Principles
EDFAs are, in general, optical amplifiers, by which the rate equation theory explains very well [6]. In order to make
the analytical model simpler, the two-state assumption and the stationary assumption are usually employed. Derived
from the two-state rate equations, the transmitting power evolution of the optical signals S  z ,   can be described
by Eq.(1). The factor  12    is the absorption cross section and  21    is the emission cross section. The factor
N tot is the population of the total ions in the erbium doped fiber. All those three factors are independent with signal
and pump, but only relative to the physical properties of the doped fiber. Only the population of the upper level N 2 ,
depends on the optical field, i.e., the pump and the signal power.
dS
  21    n2   12    n1  S   21      12     n2   12    ntot  S
dz 
     n2  ntot 12     S (1)
 S  L,   
log G     log    N tot 12     N 2    
 S  0,   
 
The spectral logarithmic gain log G    , which is the logarithm of the ratio of the output spectrum S  L,   and
the input spectrum S  0,   , can be written as the linear combination of two arbitrary spectra Ga , Gb . The
combination coefficients are related to the upper-level population  N a , N b  of each operating state. The gain

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on July 04,2022 at 21:15:22 UTC from IEEE Xplore. Restrictions apply.
 Asia Communications and Photonics Conference 2021 © Optica
T4A.98
Publishing Group 2021

spectrum Gx of an arbitrary state N x can be written as Eq.(2). Finally, the spectral gain is linearly determined by a
certain saturating coefficient k , which is highly related to the signal and pump power.
Nb  N x N  Na
Gx   Ga  x  Gb  k  Ga  1  k   Gb
Nb  N a Nb  N a (2)
 G 0  k  G1
In order to increase the precision of the measurement of Ga , Gb , principal component analysis (PCA) can be
used. For a group of measurement spectral vectors G : G1 , G2 , , Gn  , the eigenvalues   i  of the empirical
covariance matrix G T G are called latent factors. After sorting by the latent factors from maximum to minimum, the
1st and 2nd eigenvectors of G T G , i.e., 1st and 2nd principal component scores, are the most confident spectral vectors
a , G
G  b . And the inhomogeneous-form base spectral vectors G 0 , G1 are easy to calculate by least square
method (LS).

GT G  G T ΛG
 (3)
The last thing needed to determine the gain spectrum is the relationship between coefficient k , input signal
power Sin and pump P . A single-carrier curve-fit function in [7] gives an accurate prediction of the relationship
between signal power and gain. The effect of pump power is considered by the proposed modified formula Eq.(4).
The parameters  ,  ,  are rectifying exponential factors. The term P0 P  is associated with the small-signal gain
under the given pump power P . The parameter  is the pump efficiency. And the parameter L is the gain ratio
when fully saturated.
1
 P Sin  ref   
G  ref    0   L (in linear scale) (4)
 P  

 P
The procedures to predict the gain spectrum list as follow. Firstly, measure a group of spectral gain in different
saturating states (different input signal power or different pump power) and calculate the family of gain spectra.
Secondly, measure a group of single-carrier gains and fit the relationship of saturated gain. Thirdly, find the exact
spectrum in the family according to the saturated gain under certain input signal and pump power.

3. Experiment Results and Discussion

The validation of the model mentioned above is carried out by experiments, as shown in Fig.1. The EDFA under test
is a pump-variable amplifier (EDFA, SHCONNECT MFAS-B-LP). A 25-GHz optical frequency comb (Comb,
OPTOCOMB WTAS-02) is used to generate WDM power spectrum. The optical spectrum analyzer (OSA,
YOKOGAWA AQ6370C) is used to detect the power spectrum of input signal and amplified signal.

Utilizar para
Memoria
Fig.1 Gain measurement experiment setup (Comb for spectral gain and laser for single-carrier saturated gain)

3.1. Spectral Gain Measurement

The wide-band WDM power signal generated by the comb decays gradually from central wavelength to the edge,
but the central carrier (at 1535 nm) is attenuated, as shown in Fig.2 (a). Gain spectra are obtained by subtracting the
envelops of input spectra from those of amplified spectra. By controlling the attenuation of VOA, a group of gain
spectra under different input conditions are measured. As shown in Fig.2 (b), after PCA processing, the experiment
gain spectra are well fitted with the linear-combination-model described by Eq.(2), with an fitness of 0.1756 dB
(RMSE). The base spectral vectors G 0 , G1 are shown in Fig.2 (c). With the input power becoming lower, the

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on July 04,2022 at 21:15:22 UTC from IEEE Xplore. Restrictions apply.
 Asia Communications and Photonics Conference 2021 © Optica
T4A.98
Publishing Group 2021

gain spectrum tends to be closer to G1 , i.e., the saturating coefficient k approaching to 1. There is an additional
note for the exceptional ripples around 1535 nm. Because of the absence of the central carrier, the envelop curve
fitting malfunctions, which results in the sharp ripples of measured spectral gain.
(a) (b)

30

20

10 RMSE:
0.1756
0

-10
20 1.58
Rela
tive 1.56
Inpu 10
t Po 1.54 [m] 10 -6
wer gth
[dB
]
0 1.52 av elen
W

(c) (d)

20

15

10 RMSE:
0.1230
5

0
-60 0.6
-40 0.4
Input P -20 0.2 ]
ower 0 p [W
[dBm] 20 0 Pum

Fig.2 (a). WDM spectrum before (blue) and after (orange) EDFA. (b). Spectral gain fitting under different input conditions.
(c). Base spectral gain vector (black bolded lines) calculated from measurement data (colored lines with ‘x’ marks).
(d). Single-carrier gain measurement data (dots) and fitting curve (colored lines) under different saturating states.

3.2. Single-Carrier Gain Measurement

The laser (at 1550 nm) and power-meters are used to obtain the saturated gain. The pump power of EDFA varies
from 20 mW to 300 mW and the input signal power varies from -36 dBm to 10 dBm. According to the modified
single-carrier curve-fit function Eq.(4), the data-fitting result shows a good accuracy (0.1230dB RMSE), as shown in
Fig.2 (d). With the input power increasing or the pump power decreasing, the saturation intensity approaches
stronger and the saturated gain becomes lower.

4. Conclusions
The wide-band gain features of EDFA can be simply described by a data-based black-box model. The model
describes the relationship between gain, input power and pump power. Experiments varies its validation and
accuracy, with a fitness of 0.1756dB RMSE of spectral gain and 0.1230dB RMSE of single-carrier saturated gain.

Acknowledgements
This work is supported by the National Key R&D Program of China 2018YFB1801205 and NSFC under
Grants 61931010.

References
[1] J. Li et al., “16 Tbit/s Real-time Unrepeated Transmission over 420 km G.654E Fiber,” in ECOC (2020), doi: 10/gmc9pb.
[2] D. Bonnedal, “Single-setup characterization of optical fiber amplifiers,” IEEE Photon. Technol. Lett., doi: 10/fgwwjq.
[3] C. R. Giles and E. Desurvire, “Modeling erbium-doped fiber amplifiers,” J. Lightwave Technol., doi: 10/bp37pb.
[4] F. Da Ros, “Machine learning-based EDFA Gain Model Generalizable to Multiple Physical Devices,” arXiv:2009.05326
[5] S. Zhu et al., “Hybrid Machine Learning EDFA Model,” in OFC (2020), doi: 10.1364/OFC.2020.T4B.4.
[6] P. C. Becker, N. A. Olsson, J. R. Simpson, Erbium-doped fiber amplifiers: fundamentals and technology. (Academic Press, 1999), Chap. 5.
[7] D. Bonnedal, “EDFA Gain Described with a Black Box Model,” in Optical Amplifiers and Their Applications (1996), paper FAW4.

Authorized licensed use limited to: Universidad de Concepcion. Downloaded on July 04,2022 at 21:15:22 UTC from IEEE Xplore. Restrictions apply.