This thesis investigates neural-network-based wavefront and beam sensing using integrated multi-aperture photonic receivers. The device employs a Mach- Zehnder interferometer (MZI) mesh with thermo-optic phase shifters to coherently combine sampled optical fields. A closed-loop control adjusts phase settings to concentrate optical power at a selected output port, producing control parameters that encode beam characteristics such as phase differences and power ratios. This research addresses the problem of inferring the original optical field parameters from the applied phase control values using machine learning. A central contribution of the work is the complete data-driven pipeline developed for this purpose: a chip-and-control simulator is implemented to generate representative device datasets, neural-network regression is selected and validated as the inversion strategy, and experimental measurements are collected and analyzed to test the approach on real silicon photonic hardware. The analysis also includes analytical bounds on sensing precision constrained by photodiode sensitivity and a quantitative assessment of performance degradation due to non-idealities, such as imperfect directional coupler splitting ratios and thermal cross-talk between phase shifters. Numerical and experimental results demonstrate that neural networks can achieve high estimation accuracy when trained on representative datasets, even in the presence of device imperfections. Under favorable simulated conditions, the method reaches sensitivities on the order of 0.01 rad and 0.01 dB, while experimental data show errors below 0.1 rad and 0.1 dB in the appropriate power regime. The thesis further proposes a scalability-oriented sensing architecture that addresses the fragmentation limits of a direct binary-tree mesh. The findings indicate the feasibility of real-time integrated beam sensing and wavefront reconstruction, with applications ranging from adaptive optical links to general optical field monitoring and hardware-accelerated photonic processing.
Questa tesi indaga l’utilizzo di reti neurali per il wavefront sensing e il beam sensing attraverso ricevitori fotonici integrati multi-apertura. Il dispositivo oggetto dello studio impiega una griglia di interferometri Mach-Zehnder (MZI) con modulatori di fase termici per combinare coerentemente campioni del campo ottico ricevuto. Regolando i parametri di fase per concentrare la potenza in una porta di uscita designata, il ricevitore genera parametri di controllo che codificano informazioni sulle caratteristiche del fascio in ingresso, quali differenze di fase e rapporti di potenza. La tesi affronta il problema inverso: inferire i parametri originali del campo ottico dai valori di controllo di fase applicati mediante il machine learning. Un contributo centrale del lavoro è lo sviluppo della pipeline completa necessaria a validare questa idea: realizzazione del simulatore del chip e del controllo, generazione e analisi dei dataset, scelta della rete neurale come strategia di inversione e verifica sperimentale su dati acquisiti da un dispositivo fotonico reale. L’analisi include inoltre limiti analitici sulla precisione del sensing imposti dalla sensibilità dei fotodiodi e una valutazione quantitativa del degrado delle prestazioni dovuto alle non idealità del dispositivo, come coefficienti di accoppiamento imperfetti negli accoppiatori direzionali e interferenza termica tra i modulatori di fase. I risultati numerici e sperimentali dimostrano che le reti neurali possono raggiungere un’elevata precisione di stima se addestrate su dataset rappresentativi, anche in presenza di imperfezioni del dispositivo. In condizioni simulate favorevoli, il metodo raggiunge sensibilità dell’ordine di 0.01 rad e 0.01 dB, mentre i dati sperimentali mostrano errori inferiori a 0.1 rad e 0.1 dB nel regime di potenza appropriato. La tesi propone inoltre un’architettura orientata alla scalabilità che supera i limiti di frammentazione osservati nella struttura binary tree. I risultati suggeriscono la fattibilità del sensing integrato in tempo reale per la ricostruzione di fronti d’onda e fasci ottici, con possibili applicazioni nei collegamenti ottici adattivi, nel monitoraggio del campo ottico e nell’elaborazione fotonica accelerata da hardware.
A Data-Driven approach to inline wavefront sensing with integrated photonics
BOIN, ANDREA
2025/2026
Abstract
This thesis investigates neural-network-based wavefront and beam sensing using integrated multi-aperture photonic receivers. The device employs a Mach- Zehnder interferometer (MZI) mesh with thermo-optic phase shifters to coherently combine sampled optical fields. A closed-loop control adjusts phase settings to concentrate optical power at a selected output port, producing control parameters that encode beam characteristics such as phase differences and power ratios. This research addresses the problem of inferring the original optical field parameters from the applied phase control values using machine learning. A central contribution of the work is the complete data-driven pipeline developed for this purpose: a chip-and-control simulator is implemented to generate representative device datasets, neural-network regression is selected and validated as the inversion strategy, and experimental measurements are collected and analyzed to test the approach on real silicon photonic hardware. The analysis also includes analytical bounds on sensing precision constrained by photodiode sensitivity and a quantitative assessment of performance degradation due to non-idealities, such as imperfect directional coupler splitting ratios and thermal cross-talk between phase shifters. Numerical and experimental results demonstrate that neural networks can achieve high estimation accuracy when trained on representative datasets, even in the presence of device imperfections. Under favorable simulated conditions, the method reaches sensitivities on the order of 0.01 rad and 0.01 dB, while experimental data show errors below 0.1 rad and 0.1 dB in the appropriate power regime. The thesis further proposes a scalability-oriented sensing architecture that addresses the fragmentation limits of a direct binary-tree mesh. The findings indicate the feasibility of real-time integrated beam sensing and wavefront reconstruction, with applications ranging from adaptive optical links to general optical field monitoring and hardware-accelerated photonic processing.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Boin_Thesis.pdf
non accessibile
Descrizione: Tesi
Dimensione
78.81 MB
Formato
Adobe PDF
|
78.81 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_Boin_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
16.31 MB
Formato
Adobe PDF
|
16.31 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/260634