Optical fiber nonlinearity compensation using neural networks
Aspects of the present disclosure describe systems, methods and structures for optical fiber nonlinearity compensation using neural networks that advantageously employ machine learning (ML) algorithms for nonlinearity compensation (NLC) that advantageously provide a system-agnostic model independent of link parameters, and yet still achieve a similar or better performance at a lower complexity as compared with prior-art methods. Systems, methods, and structures according to aspects of the present disclosure include a data-driven model using the neural network (NN) to predict received signal nonlinearity without prior knowledge of the link parameters. Operationally, the NN is provided with intra-channel cross-phase modulation (IXPM) and intra-channel four-wave mixing (IFWM) triplets that advantageously provide a more direct pathway to underlying nonlinear interactions.
1. An improved nonlinearity compensation (NLC) method for optical transmission networks employing a neural network (NN-NLC) said improved NN-NLC method CHARACTERIZED IN THAT:
the NN-NLC is performed without a knowledge of transmission link parameters;
said NN fed with intra-channel cross-phase modulation (IXPM) and intra-channel four-wave mixing (IFWM) triplets, said triplets selected according to the following criteria:
n
≤
min
{
ρ
⌈
L
/
2
⌉
m
,
⌈
L
2
⌉
}
where ρ is a scaling factor to determine the width of triplets at edges
(
m
=
±
⌈
L
2
⌉
)
,
L is symbol window length, ┌⋅┐ and |⋅| are rounding upper towards nearest integer and absolute operation, respectively.
2. The improved method of claim 1 FURTHER CHARACTERIZED IN THAT:
the unknown link parameters are ones selected from the group consisting of: optical dispersion, fiber nonlinearity and span length.
3. The improved method of claim 2 FURTHER CHARACTERIZED IN THAT:
the optical transmission network is a software-defined mesh network.
4. The improved method of claim 3 FURTHER CHARACTERIZED IN THAT:
the NN-NLC includes both training and execution stages.
5. The improved method of claim 4 FURTHER CHARACTERIZED IN THAT:
the NN-NLC execution stage is performed only at a transmitting side of a transmission link.
6. The improved method of claim 5 FURTHER CHARACTERIZED IN THAT:
the NN-NLC training stage is performed operating on soft data obtained from carrier phase recovery in a receiver's digital signal processor (DSP).
7. The improved method of claim 6 FURTHER CHARACTERIZED IN THAT:
the NN-NLC training stage checks performance of a neural network model against a cross-validation (CV) dataset to generate a learned model.
8. The improved method of claim 6 FURTHER CHARACTERIZED IN THAT:
the learned model is applied to data for all channel powers in the execution stage.
9. An improved nonlinearity compensation method for optical networks employing a neural network having training and execution stages said improved method comprising:
during the training stage, generating a neural network model by operating on soft data from carrier phase recovery in a receiver digital signal processor;
evaluating performance of the model with a cross-validation dataset to generate a learned model;
training the learned model by transmitting known but randomly generated patterns and searching for a best node tensor parameter that minimize mean square error (MSE) between transmitted and received symbols given by:
MSE
=
1
B
∑
i
=
1
B
H
i
-
(
H
^
i
-
H
^
i
,
NL
)
2
,
where Ĥ i and Ĥ i,NL and respectively, are the received symbols and estimated nonlinearity for polarization H (pol-H), H i is the transmitted symbol, and |.| is an absolute operation;
applying intra-channel cross-phase modulation (IXPM) and intra-channel four-wave mixing (IFWM) triplets to the neural network, to estimate the nonlinearity, without prior knowledge of transmission link parameters;
said triplets selected according to the following criteria:
n
≤
min
{
ρ
⌈
L
/
2
⌉
m
,
⌈
L
2
⌉
}
where ρ is a scaling factor to determine the width of triplets at edges
(
m
=
±
⌈
L
2
⌉
)
,
L is symbol window length, ┌⋅┐ and |⋅| are rounding upper towards nearest integer and absolute operation, respectively.
10. The improved nonlinearity compensation method of claim 9 wherein the neural network model generated during the training stage is applied to one of a transmitter and the receiver.