oopt-gnpy/docs/model.rst

.. _physical-model:

Physical Model used in GNPy
===========================

QoT-E including ASE noise and NLI accumulation 
----------------------------------------------

The operations of PSE simulative framework are based on the capability to
estimate the QoT of one or more channels operating lightpaths over a given
network route. For backbone transport networks, we can suppose that
transceivers are operating polarization-division-multiplexed multilevel
modulation formats with DSP-based coherent receivers, including equalization.
For the optical links, we focus on state-of-the-art amplified and uncompensated
fiber links, connecting network nodes including ROADMs, where add and drop
operations on data traffic are performed. In such a transmission scenario, it
is well accepted
:cite:`vacondio_nonlinear_2012,bononi_modeling_2012,carena_modeling_2012,mecozzi_nonlinear_2012,secondini_analytical_2012,johannisson_perturbation_2013,dar_properties_2013,serena_alternative_2013,secondini_achievable_2013,poggiolini_gn-model_2014,dar_accumulation_2014,poggiolini_analytical_2011,savory_approximations_2013,bononi_single-_2013,johannisson_modeling_2014`
to assume that transmission performances are limited by the amplified
spontaneous emission (ASE) noise generated by optical amplifiers and and
by nonlinear propagation effects: accumulation of a Gaussian disturbance
defined as nonlinear interference (NLI) and generation of phase noise.
State-of-the-art DSP in commercial transceivers are typically able to
compensate for most of the phase noise through carrier-phase estimator
(CPE) algorithms, for modulation formats with cardinality up to 16, per
polarization state
:cite:`poggiolini_recent_2017,schmidt_experimental_2015,fehenberger_experimental_2016`.
So, for backbone networks covering medium-to-wide geographical areas, we
can suppose that propagation is limited by the accumulation of two
Gaussian disturbances: the ASE noise and the NLI. Additional impairments
such as filtering effects introduced by ROADMs can be considered as
additional equivalent power penalties depending on the ratio between the
channel bandwidth and the ROADMs filters and the number of traversed
ROADMs (hops) of the route under analysis. Modeling the two major
sources of impairments as Gaussian disturbances, and being the receivers
*coherent*, the unique QoT parameter determining the bit error rate
(BER) for the considered transmission scenario is the generalized
signal-to-noise ratio (SNR) defined as

.. math::

   {\text{SNR}}= L_F \frac{P_{\text{ch}}}{P_{\text{ASE}}+P_{\text{NLI}}} = L_F \left(\frac{1}{{\text{SNR}}_{\text{LIN}}}+\frac{1}{{\text{SNR}}_{\text{NL}}}\right)^{-1}


where :math:`P_{\text{ch}}` is the channel power,
:math:`P_{\text{ASE}}` and :math:`P_{\text{NLI}}` are the power levels of the disturbances 
in the channel bandwidth for ASE noise and NLI, respectively.
:math:`L_F` is a parameter assuming values smaller or equal than one
that summarizes the equivalent power penalty loss such as 
filtering effects. Note that for state-of-the art equipment, filtering
effects can be typically neglected over routes with few hops
:cite:`rahman_mitigation_2014,foggi_overcoming_2015`.

To properly estimate :math:`P_{\text{ch}}` and :math:`P_{\text{ASE}}`
the transmitted power at the beginning of the considered route must be
known, and losses and amplifiers gain and noise figure, including their
variation with frequency, must be characterized. So, the evaluation of
:math:`{\text{SNR}}_{\text{LIN}}` *just* requires an accurate
knowledge of equipment, which is not a trivial aspect, but it is not
related to physical-model issues. For the evaluation of the NLI, several
models have been proposed and validated in the technical literature
:cite:`vacondio_nonlinear_2012,bononi_modeling_2012,carena_modeling_2012,mecozzi_nonlinear_2012,secondini_analytical_2012,johannisson_perturbation_2013,dar_properties_2013,serena_alternative_2013,secondini_achievable_2013,poggiolini_gn-model_2014,dar_accumulation_2014,poggiolini_analytical_2011,savory_approximations_2013,bononi_single-_2013,johannisson_modeling_2014`.
The decision about which model to test within the PSE activities was
driven by requirements of the entire PSE framework:

i. the model must be *local*, i.e., related individually to each network
element (i.e. fiber span) generating NLI, independently of preceding and
subsequent elements; and ii. the related computational time must be compatible
with interactive operations. 

So, the choice fell on the Gaussian Noise
(GN) model with incoherent accumulation of NLI over fiber spans
:cite:`poggiolini_gn-model_2014`. We implemented both the
exact GN-model evaluation of NLI based on a double integral (Eq. (11) of
:cite:`poggiolini_gn-model_2014`) and its analytical
approximation (Eq. (120-121) of
:cite:`poggiolini_analytical_2011`). We performed several
validation analyses comparing results of the two implementations with
split-step simulations over wide bandwidths
:cite:`pilori_ffss_2017`, and results clearly showed that
for fiber types with chromatic dispersion roughly larger than 4
ps/nm/km, the analytical approximation ensures an excellent accuracy
with a computational time compatible with real-time operations.

The Gaussian Noise Model to evaluate the NLI
--------------------------------------------

As previously stated, fiber propagation of multilevel modulation formats
relying on the polarization-division-multiplexing  generates impairments that
can be summarized as  a disturbance called nonlinear interference (NLI), when
exploiting a DSP-based coherent receiver, as in all state-of-the-art equipment.
From a practical point of view, the NLI can be modeled as an additive Gaussian
random process added by each fiber span, and whose strength depends on the cube
of the input power spectral density and on the fiber-span parameters. 

Since the introduction in the market in 2007 of the first transponder based on
such a transmission technique, the scientific community has intensively worked
to define the propagation behavior of such a trasnmission technique.  First,
the role of in-line chromatic dispersion compensation has been investigated,
deducing that besides being not essential, it is indeed detrimental for
performances :cite:`curri_dispersion_2008`.  Then, it has been observed that
the fiber propagation impairments are practically summarized by the sole NLI,
being all the other phenomena compensated for by the blind equalizer
implemented in the receiver DSP :cite:`carena_statistical_2010`.  Once these
assessments have been accepted by the community, several prestigious research
groups have started to work on deriving analytical models able to estimating
the NLI accumulation, and consequentially the generalized SNR that sets the
BER, according to the transponder BER vs. SNR performance.  Many models
delivering different levels of accuracy have been developed and validated. As
previously clarified, for the purposes of the PSE framework, the  GN-model with
incoherent accumulation of NLI over fiber spans has been selected as adequate.
The reason for such a choice is first such a model being a "local" model, so
related to each fiber spans, independently of the preceding and succeeding
network elements. The other model characteristic driving the choice is the
availability of a closed form for the model, so permitting a real-time
evaluation, as required by the PSE framework.  For a detailed derivation of the
model, please refer to :cite:`poggiolini_analytical_2011`, while a qualitative
description can be summarized as in the following.  The GN-model assumes that
the channel comb propagating in the fiber is well approximated by unpolarized
spectrally shaped Gaussian noise. In such a scenario, supposing to rely - as in
state-of-the-art equipment - on a receiver entirely compensating for linear
propagation effects, propagation in the fiber only excites the four-wave mixing
(FWM) process among the continuity of the tones occupying the bandwidth. Such a
FWM generates an unpolarized complex Gaussian disturbance in each spectral slot
that can be easily evaluated extending the FWM theory from a set of discrete
tones - the standard FWM theory introduced back in the 90s by Inoue
:cite:`Innoue-FWM`- to a continuity of tones, possibly spectrally shaped.
Signals propagating in the fiber are not equivalent to Gaussian noise, but
thanks to the absence of in-line compensation for choromatic dispersion, the
become so, over short distances.  So, the Gaussian noise model with incoherent
accumulation of NLI has estensively proved to be a quick yet accurate and
conservative tool to estimate propagation impairments of fiber propagation.
Note that the GN-model has not been derived with the aim of an *exact*
performance estimation, but to pursue a conservative performance prediction.
So, considering these characteristics, and the fact that the NLI is always a
secondary effect with respect to the ASE noise accumulation, and - most
importantly - that typically linear propagation parameters (losses, gains and
noise figures) are known within a variation range, a QoT estimator based on the
GN model is adequate to deliver performance predictions in terms of a
reasonable SNR range, rather than an exact value.  As final remark, it must be
clarified that the GN-model is adequate to be used when relying on a relatively
narrow bandwidth up to few THz. When exceeding such a bandwidth occupation, the
GN-model must be generalized introducing the interaction with the Stimulated
Raman Scattering in order to give a proper estimation for all channels
:cite:`cantono2018modeling`.  This will be the main upgrade required within the
PSE framework.

.. bibliography:: biblio.bib