8 November 2019 in Journal of Lightwave Technology
A wide range of highly-relevant problems in programmable and integrated photonics, optical amplification, and communication deal with inverse system design. Typically, a desired output (usually a gain profile, a noise profile, a transfer function or a similar continuous function) is given and the goal is to determine the corresponding set of input parameters (usually a set of input voltages, currents, powers, and wavelengths). We present a novel method for inverse system design using machine learning and apply it to Raman amplifier design. Inverse system design for Raman amplifiers consists of selecting pump powers and wavelengths that would result in a targeted gain profile. This is a challenging task due to highly-complex interaction between pumps and Raman gain. Using the proposed framework, highly-accurate predictions of the pumping setup for arbitrary Raman gain profiles are demonstrated numerically in C and C+L-band, as well as experimentally in C band, for the first time. A low mean (0.46 and 0.35 dB) and standard deviation (0.20 and 0.17 dB) of the maximum error are obtained for numerical (C+L-band) and experimental (C-band) results, respectively, when employing 4 pumps and 100 km span length. The presented framework is general and can be applied to other inverse problems in optical communication and photonics in general.
26 October 2020 in Journal of Lightwave Technology
Optical communication systems, operating in C-band, are reaching their theoretically achievable capacity limits. An attractive and economically viable solution to satisfy the future data rate demands is to employ the transmission across the full low-loss spectrum encompassing O, E, S, C, and L band of the single mode fibers (SMF). Utilizing all five bands offers a bandwidth of up to ~53.5 THz (365 nm) with loss below 0.4 dB/km. A key component in realizing multi-band optical communication systems is the optical amplifier. Apart from having an ultra-wide gain profile, the ability of providing arbitrary gain profiles, in a controlled way, will become an essential feature. The latter will allow for signal power spectrum shaping which has a broad range of applications such as the maximization of the achievable information rate x distance product, the elimination of static and lossy gain flattening filters (GFF) enabling a power efficient system design, and the gain equalization of optical frequency combs. In this paper, we experimentally demonstrate a multi-band (S+C+L) programmable gain optical amplifier using only Raman effects and machine learning. The amplifier achieves >1000 programmable gain profiles within the range 3.5 to 30 dB, in an ultra-fast way and a very low maximum error of 1.6 · 10 -2 dB/THz over an ultra-wide bandwidth of 17.6-THz (140.7-nm).
6 August 2020 in Journal of Lightwave Technology
An ultra-fast machine learning based method for accurate predictions of gain and amplified spontaneous emission (ASE) noise profiles of Raman amplifiers is introduced. It is an alternative to high-complexity and time-consuming standard approaches, which are based on the numerical solution of sets of nonlinear differential equations. Main relevance resides on its possible application in real-time network controllers for future multi-band optical line systems where Raman amplification will be required to cope with capacities beyond the standard C-band. Here we consider as an example the C+L-band scenario with different input load conditions: full load and partial loads. For the case of full load it has been recently shown a neural network (NN) capable of highly accurate predictions. Real optical networks are not usually operated only in full load conditions: the load can dynamically vary over time and the behavior of the Raman amplifier depends on it. In this paper we introduce a new NN model and we show its higher accuracy when the line system is not fully loaded: we define it as the load aware neural network. Applying this new approach we can predict both gain and ASE noise profiles in Raman amplifiers with high accuracy under any load conditions: we demonstrate almost 100% of maximum prediction errors to be lower than 0.5 dB.
31 October 2022 in Journal of Lightwave Technology
The problem of Raman amplifier optimization is studied. A differentiable interpolation function is obtained for the Raman gain coefficient using machine learning (ML), which allows for the gradient descent optimization of forward-propagating Raman pumps. Both the frequency and power of an arbitrary number of pumps in a forward pumping configuration are then optimized for an arbitrary data channel load and span length. The forward propagation model is combined with an experimentally-trained ML model of a backward-pumping Raman amplifier to jointly optimize the frequency and power of the forward amplifier's pumps and the powers of the backward amplifier's pumps. The joint forward and backward amplifier optimization is demonstrated for an unrepeatered transmission of 250 km. A gain flatness of $<$ 1 dB over 4 THz is achieved. The optimized amplifiers are validated using a numerical simulator.
27 January 2016 in Journal of Lightwave Technology
Our recently proposed cognitive methodology for optical amplifier gain adjustment, that relies on case-based reasoning, showed optical signal-to-noise ratio improvements over time demonstrating the cognition process regardless the deployed amplifier type. In this paper, we extend our preliminary analysis exploring the cognitive methodology benefits for different and larger network topologies. The obtained results show agreement between networks, demonstrating the methodology suitability regardless the network scenario.
29 September 2022 in Journal of Lightwave Technology
In this paper, we show that by combining experimental data from different optical fibers, we can build a fiber-agnostic neural-network to model the Raman amplifier. The fiber-agnostic NN model can predict the gain profile of a new fiber type (unseen by the model during training) with a maximum absolute error as low as 0.22 dB. We show that this generalization is only possible when the unseen fiber parameters are similar to the fibers used to build the model. Therefore, a training dataset with a wide range of optical fibers parameters is needed to enhance the chance of accurately predicting the gain of a new fiber. This implies that time-consuming experimental measurements of various fiber types can be avoided. For that, here we extend and improve our general model by numerically generating the dataset. By doing so, it is possible to generate uniformly distributed data that covers a wide range of optical fiber types. The results show that the averaged maximum prediction error is reduced when compared to the limited experimental data-based general models. As the second and final contribution of this work, we propose the use of transfer learning (TL) to re-train the numerical data-based general model using just a few experimental measurements. Compared with the fiber-specific models, this TL-upgraded general model reaches very similar accuracy, with just 3.6% of the experimental data . These results demonstrate that the already fast and accurate NN-based RA models can be upgraded to have strong fiber generalization capabilities.