Research activities in optics and photonics are applied to the digital sector in several areas: information transmission, data processing, and unconventional imaging.
Information transmission and machine learning
The first research priority focuses on information transmission and data processing using machine learning algorithms.
The main challenges facing future optical networks are increasing data rates and reducing energy consumption. To address these challenges, the optical layer must become agile, resilient, and intelligent to respond to unpredictable data traffic demands while optimizing resource allocation. In the context of implementing and evaluating next-generation optical networks, an initial research activity was conducted in collaboration with ETS Montréal to study the coexistence of 10G and 100G data rates in optical networks. Simulations and experimental validation were conducted to assess the impairments affecting the 100G channel.
Another approach that we have been exploring for several years involves developing a software tool capable of both accurately modeling the optical physical layer and creating and implementing the abstractions required to integrate and test new optical functionalities, such as optical packet switching, and QKD, and assessing their impact on network-layer performance (QoT, blocking probability, energy consumption, etc.).
We have also investigated the combined benefits of machine learning techniques and photonic techniques used in optical transmission systems. We initially implemented these techniques to study the impact of
nonlinear effects in optical fibers on transmission performance. To this end, we proposed a technique for optimizing signal processing parameters to mitigate these effects [Lehmann and Frignac, 2018].
We have also investigated machine learning techniques for the dynamic control of optical power variation across different channels amplified by EDFAs (Erbium-Doped Fiber Amplifiers). Since physical-layer impairments accumulate along the optical path, we correlated the optical signal-to-noise ratio (OSNR) and bit error rate (BER) with power excursion to estimate the quality of transmission (QoT) [Freire-Hermelo et al., 2021b]. This work continues through an ongoing collaboration with ETS Montréal.
Machine learning techniques have also been implemented to predict network equipment failures. The goal is to significantly improve quality of service and reduce costs by proactively maintaining network equipment. As part of a CIFRE PhD project with SPIE ICS, a comprehensive review of the latest developments in network fault prediction was conducted. The review focused on identifying machine learning methods that could address this challenge [Murphy et al., 2023]. We compared decision-tree-based machine learning methods using a real-world dataset generated from SPIE ICS's monitoring of optical access networks. A new cost-estimation metric was proposed.
Neuromorphic photonic systems for data processing
In 2018, we launched a new research initiative dedicated to neuromorphic photonic systems based on the concept of reservoir computing. Photonics systems are particularly well-suited for physical implementations of these neuromorphic systems because of their simple learning process, the very high speed of optical components, their potential for parallel signal processing, and the wide range of available nonlinearities, and their very low energy consumption compared to AI systems emulated on conventional computers. First, we conducted numerical studies, then we developed several physical implementations based on optoelectronic devices for applications including speech recognition, data classification, and, more specifically, the early detection of neurodegenerative diseases. For the latter, we were able to demonstrate state-of-the-art classification performance, with significant improvements in computational and energy efficiency compared to other methods used for this application [Mwamsojo et al., 2022b, Mwamsojo et al., 2022a].
Photonics components for optical information transmission and processing
The second research priority involves activities centered on photonic components. Developing innovative photonic devices is essential for both information transmission and data processing. To address this need, we are researching and developing reconfigurable pulsed laser sources and III-V semiconductor laser sources.
Reconfigurable pulsed laser sources
Building on our previous work on wavelength-discriminating devices, we have leveraged these structures to explore applications such as laser pulse manufacturing and pulse repetition-rate multiplication. We have developed a new synthesis method for manufacturing time pulses based on cascading birefringent cells. The cells’ optogeometric parameters were determined using a Particle Swarm Optimization (PSO) algorithm. We generated symmetric and asymmetric pulse profiles, thereby demonstrating the feasibility of the proposed approach in the picosecond and femtosecond ranges [Halassi et al., 2020; Driouche et al., 2022a]. These anisotropic structures, which exhibit periodic quadratic phases, have also made it possible to demonstrate the feasibility of a pulse repetition rate multiplier based on the temporal Talbot effect. The filter was synthesized using a genetic algorithm. With an input repetition rate of 10 GHz, pulse trains with repetition rates multiplied by a given factor were generated without altering the pulse characteristics
[Driouche et al., 2022b, Driouche et al., 2022c].
On-chip optical frequency combs
We are also developing optical sources to meet the growing demand for higher data rates while keeping energy consumption low and limiting the environmental footprint of the
ICT sector, which has grown in recent years. On-chip optical frequency combs have the potential to revolutionize massively parallel, ultra-high-speed optical communications through wavelength division multiplexing (WDM). They can also contribute to the development of integrated neuromorphic photonic chips. This recent work is part of two ongoing projects: ANR ORCHID and ANR HybridCombs. As part of these projects, we are developing a new class of chip-scale optical frequency comb sources that combines the distinct advantages of soliton Kerr microcombs with those of mode-locked lasers based on nanostructures made from III-V semiconductor materials. This optical technology will ultimately enable the generation of a multi-wavelength optical frequency comb spanning several THz, allowing for optical data transmission at rates of several tens of terabits per second using a single component [Kemal et al., 2020; Merghem, 2023b; Verolet et al., 2020].
Dynamics of semiconductor lasers in the presence of external disturbances
This work builds on several years of research into the emission properties of single-section Fabry–Pérot lasers in the presence of external optical feedback. More specifically, we examine how lasers behave when subjected to disturbances of various physical origins, such as optical injection, electrical feedback, optical feedback, and optoelectronic feedback. Our goal is to develop novel approaches that can be adapted to a range of applications, including coherent transmission, optical chaos encryption, LiDAR, microwave sources, and signal processing. This study models the operating regime of a laser system (monomodal, periodic, or chaotic) using several classical mathematical tools, such as Feigenbaum, Hopf, and saddle-node bifurcations, as well as the Lyapunov exponent and Lyapunov fractal [Zou et al., 2020].
Computational imaging for healthcare and life sciences
The final research priority focuses on unconventional imaging. This work began in 2017, with the development of microscopic imaging systems to observe and analyze biological samples. The systems studied are based on a computational imaging approach. More specifically, the imaging systems developed produce spatial maps of optical parameters measured in the sample. These parameters are obtained by solving an inverse problem based on a model of light–matter interaction (e.g. absorption and dipole interaction) combined with an image-formation model.
The experimental approach adopted is Fourier ptychographic microscopy (FPM). FPM produces bimodal images (intensity and phase) with a high super-resolution factor (up to 6x) and enables spatial resolutions beyond the limits of conventional optical microscopy (a resolution of 100 nm has been achieved).
The first challenge is addressed through the co-design of the sensor and image reconstruction algorithms. Regarding the sensor, a study was proposed to examine the efficiency of the sample excitation scheme (patent FR2108256). From an algorithmic perspective, the work introduced original hybrid methods combining deep learning and the optimization of deep networks that explicitly incorporate physical equations. This expanded the rules for designing the FPM microscope under study [Bouchama et al., 2023a] and enabled access to a wider range of sample characteristics, including vectorial properties [Baroni et al., 2022]. Finally, a technique for improving image quality based on autofocus was introduced [Bouchama et al., 2023b; Bouchama et al., 2022]. Overall, this work enables the virtualization of samples, which can then be digitally manipulated by an expert for observation under various experimental conditions, creating a digital twin of the sample.
The second challenge focuses on developing high-performance automated diagnostic systems for blood diseases in hematology (e.g. leukemia) and parasitology (e.g. malaria). We demonstrated that exploiting the bimodal nature of the images could significantly improve the detection and classification of pathological cells [Hassini et al., 2023]. Fully automated malaria diagnosis from blood smears now appears to be within reach. This work is supported by three industrial collaborations, including a strategic partnership between Telecom SudParis and TRIBVN, an ANR-LabCom OT4D project, and a Bpifrance HDH-TAMIS project. There are also several academic collaborations with Avicenne Hospital, Saint-Antoine Hospital, Pitié-Salpêtrière Hospital, Paris-Sorbonne University, Généthon, and Inserm.
