In this thesis work we study how to characterise squeezed light produced by spontaneous four-wave mixing (SFWM) in a microring resonator using a machine-learning-based approach. In the introductory chapter, we give a theoretical description of squeezed states of light within a framework which is independent of the structure of the system, introducing some definitions about entanglement and defining the Schmidt number. We presented some strategies to generate squeezed states of light, pioneering experiments to integrated structures such as ring resonators. We give particular attention to the description of SFWM in this system, by deriving the complete Hamiltonian for dual-pump and single-pump SFWM in a microring resonator. We describe the dynamics of the system and the limits of an analytic approach. We briefly introduce yellowsubmaring, a simulation Python tool provided by Xanadu Quantum Technologies, along with some elements of machine learning. In the second chapter of the thesis we present the results of our study of squeezed light generated by single-pump SFWM. We specified the most important equations describing the dynamics of the field operators involved in this process and used yellowsubmaring to generate several training sets for our machine learning analysis. We constructed the training sets by varying the parameters of the simulation and studying different regimes for the Schmidt number and the mean number of the geerated photons. We used the sets to train a kernel SVR to estimate the Schmidt number, reporting the results in the form of R-squared score, mean error and training time, together with the learning curves. In the third chapter, we consider dual-pump SFWM. Again, we generated the training sets and trained our algorithm. Our main goal was to understand whether it was possible to exploit machine learning to reduce the number of measurements necessary to estimate the separability of the state. We did this by approaching the task as a problem of image recognition (which is indeed a task where the use of machine learning is common), since yellowsubarming can provide us with the images of the real and imaginary part of the density matrix of the state, also experimentally accessible by measuring the opportune quadrature operators. We construct several training sets by varying the pump energy and duration, each of which with lower-resolution images, computing the Schmidt number for each element. These data sets were used to train a kernel SVR, and we reported the results in the form of R-squared score and mean error. We performed a principal component analysis on the data sets to reduce the effective size of the problem and look for alternative features as input for the regression model. Based on the achieved results for single-pump and dual-pump SFWM, we can conclude that our algorithm could estimate the Schmidt number with enough accuracy. Finally, we discussed a possible comparison between the Schmidt decomposition and the principal component analysis. In this thesis work we showed that it is possible to leverage machine learning techniques, along with yellowsubmaring simulation tool, to estimate the separability of light generated by SFWM in a microring resonator. The achieved results suggest that the trained algorithm is reliable in estimating the Schmidt number with enough accuracy, so that a reduction of the number of required experimental measurements seems feasible. The ability to model and add experimental-like noise to the simulated data used to train the machine learning algorithm is a possible future development, as well as considering different systems or pumping schemes, such as the double-pulse configuration.
In questo lavoro di tesi studiamo come caratterizzare la luce squeezed prodotta tramite spontaneous four-wave mixing (SFWM) in un risuonatore ad anello usando un approccio basato sul machine learning. Nel capitolo introduttivo forniamo una descrizione teorica degli stati squeezed della luce in un framework indipendente della struttura del sistema, proseguendo con alcune definizioni riguardo all’entanglement e alla definizione dello Schmidt number. Presentiamo alcune strategie per generare tali stati, dai primi esperimenti fino a strutture integrate come il risuonatore ad anello. Descriviamo lo SFWM in questo sistema, derivando l’Hamiltoniana completa per i processi dual-pump e single-pump SFWM in un risuonatore ad anello. Descriviamo la dinamica del sistema e i limiti dell’approccio analitico. Introduciamo brevemente yellowsubmaring, una libreria Python per la simulazione fornita da Xanadu Quantum Technologies, insieme ad elementi di machine learning. Nel secondo capitolo della tesi presentiamo i risultati del nostro studio sulla luce squeezed generata tramite single-pump SFWM. Ricaviamo le principali equazioni che descrivono la dinamica degli operatori di campo coinvolti nel processo e usiamo yellowsubmaring per generare diversi training set per l’analisi basata su machine learning. Costruiamo i training set variando i parametri della simulazione e studiando diversi regimi di Schmidt number e numero medio di fotoni generati. Usiamo poi i set per allenare un kernel SVR per stimare lo Schmidt umber, riportando i risultati tramite lo score R-quadro, l’errore medio e il tempo di training, insieme alle curve di apprendimento. Nel terzo capitolo consideriamo il dual-pump SFWM. Generiamo i training set e alleniamo l’algoritmo, Troviamo delle correlazioni tra lo Schmidt number e i parametri di input della simulazione, in particolare l’energia e la durata della pompa. Il nostro obiettivo primario è capire se sia possibile sfruttare il machine learning per ridurre il numero di misure necessarie per stimare la separabilità dello stato. Questo viene fatto affrontando il problema in termini di riconoscimento di immagini (compito per cui l’uso del machine learning è pratica comune), dal momento che yellowsubmaring ci può fornire le immagini della parte reale ed immaginaria della matrice densità dello stato, accessibile sperimentalmente misurando le opportune quadrature. Costruiamo poi diversi training set variando la durata e l’energia della pompa, con immagini a risoluzione più bassa, calcolando lo Schmidt number per ciascun elemento. Questi data set vengono usati per allenare un kernel SVR, di cui riportiamo i risultati usando ancora uno score R-quadro e l’errore medio. Facciamo poi un’analisi delle componenti principali dei data set per ridurre la dimensione reale del problema e cercare delle feature alternative da usare come input per il modello di regressione. Basandosi sui risultati ottenuti, possiamo concludere che il nostro algoritmo sia in grado di stimare lo Schmidt number con sufficiente accuratezza. Discutiamo poi un possibile confronto tra la Schmidt decomposition e l’analisi delle componenti principali. In conclusione, abbiamo quindi mostrato come sia possibile sfruttare in modo opportuno tecniche di machine learning per stimare la separabilità della luce generata tramite SFWM in in risuonatore ad anello. I risultati ottenuti indicano che l’algoritmo sia affidabile nella stima dello Schmidt number con sufficiente accuratezza, così che una riduzione del numero di misure sperimentali richieste sembra possibile. La capacità di modellizzare e aggiungere un rumore, compatibile con quello rilevato sperimentalmente, ai dati usati per allenare l’algoritmo di machine learning è un possibile sviluppo futuro di questo lavoro, così come considerare sistemi o pompe diverse, come la configurazione basata su doppi impulsi.
Caratterizzazione di Luce Squeezed da un Microrisuonatore ad Anello tramite Machine Learning
MARAGNANO, DIEGO
2021/2022
Abstract
In this thesis work we study how to characterise squeezed light produced by spontaneous four-wave mixing (SFWM) in a microring resonator using a machine-learning-based approach. In the introductory chapter, we give a theoretical description of squeezed states of light within a framework which is independent of the structure of the system, introducing some definitions about entanglement and defining the Schmidt number. We presented some strategies to generate squeezed states of light, pioneering experiments to integrated structures such as ring resonators. We give particular attention to the description of SFWM in this system, by deriving the complete Hamiltonian for dual-pump and single-pump SFWM in a microring resonator. We describe the dynamics of the system and the limits of an analytic approach. We briefly introduce yellowsubmaring, a simulation Python tool provided by Xanadu Quantum Technologies, along with some elements of machine learning. In the second chapter of the thesis we present the results of our study of squeezed light generated by single-pump SFWM. We specified the most important equations describing the dynamics of the field operators involved in this process and used yellowsubmaring to generate several training sets for our machine learning analysis. We constructed the training sets by varying the parameters of the simulation and studying different regimes for the Schmidt number and the mean number of the geerated photons. We used the sets to train a kernel SVR to estimate the Schmidt number, reporting the results in the form of R-squared score, mean error and training time, together with the learning curves. In the third chapter, we consider dual-pump SFWM. Again, we generated the training sets and trained our algorithm. Our main goal was to understand whether it was possible to exploit machine learning to reduce the number of measurements necessary to estimate the separability of the state. We did this by approaching the task as a problem of image recognition (which is indeed a task where the use of machine learning is common), since yellowsubarming can provide us with the images of the real and imaginary part of the density matrix of the state, also experimentally accessible by measuring the opportune quadrature operators. We construct several training sets by varying the pump energy and duration, each of which with lower-resolution images, computing the Schmidt number for each element. These data sets were used to train a kernel SVR, and we reported the results in the form of R-squared score and mean error. We performed a principal component analysis on the data sets to reduce the effective size of the problem and look for alternative features as input for the regression model. Based on the achieved results for single-pump and dual-pump SFWM, we can conclude that our algorithm could estimate the Schmidt number with enough accuracy. Finally, we discussed a possible comparison between the Schmidt decomposition and the principal component analysis. In this thesis work we showed that it is possible to leverage machine learning techniques, along with yellowsubmaring simulation tool, to estimate the separability of light generated by SFWM in a microring resonator. The achieved results suggest that the trained algorithm is reliable in estimating the Schmidt number with enough accuracy, so that a reduction of the number of required experimental measurements seems feasible. The ability to model and add experimental-like noise to the simulated data used to train the machine learning algorithm is a possible future development, as well as considering different systems or pumping schemes, such as the double-pulse configuration.È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/14530