PRIN 2022 - TREASURE

This project is aimed at developing a new paradigm of sensing by combining complex systems, machine learning and sensing and leveraging on the rich nonlinear dynamics of optical resonators, which can enable measuring temporal variations while increasing precision. The system learns environmental fluctuations considered as noise and removes them, part of the intelligence being inherent to the sensor itself and enabling computations of a subset of the post-processing, hence enhancing the dynamic capacity and energy efficiency. A possible implementation of the sensor to decorrelate local from global perturbations is depicted in Fig. 4. Since only part of the dynamic system is exposed to local perturbations, the dynamics will be affected in a different way when global perturbations are present.

The activities done so far include the testing of the system in a Distributed Acoustic Sensor (DAS), involving vibration-induced changes in Rayleigh Backscattering (RBS) intensity as shown Fig. 5. The signal from a narrow linewidth laser is used to generate coherent pulses which were amplified, filtered and sent into the fiber through the circulator, which collects the RBS in its return port. The response was detected and reproduced using an arbitrary waveform generator (500 MHz bandwidth, 2GB memory) connected to an optical modulator as shown in Fig 9 (a). The modulated signal was then fed to the silicon-on-insulator photonic integrated circuit (PIC, see Fig. 9 b), consisting of a network of silicon micro-ring resonators (MRRs). 

This system is employed as an untrained photonic neural network, following the reservoir computing (RC) paradigm. In particular, we excited the PIC with the RBS of each laser pulse inserted into the fiber sensor, each followed by constant optical power lasting 10 µs supporting the self-pulsing response of the photonic network and a low-power pause to reset the network memory. Experimental results show that, for certain input laser wavelengths and power, the self-pulsing response of the network could well separate different states of the actuator applied to the fiber sensor (Fig. 10). Moreover, we applied a linear regression to a down-sampled version of the network response, which could effectively learn to retrieve the fiber oscillation state with a coefficient of determination up to 0.97.

ENTE PROMOTORE: MUR
NOME PROGETTO: TREASURE
PERIODO E DURATA: 24 mesi, 28/09/2023 – 27/09/2025
FINANZIAMENTO: € 97.535
COORDINATORI: Scuola Superiore Sant’Anna di Pisa
REFERENTI SSSA: Yonas Seifu Muanenda