Exploiting light interference and modulation, programmable optical processors (POPs) allow the implementation of computations directly in the analog optical domain with on-premise, real-time reconfiguration of the behavior of the optical component. Through Silicon Photonics (SiPh), POPs take advantage of the well-consolidated manufacturing process used for electronic integrated circuits to fabricate optical components directly on silicon wafers, enabling cost-effective miniaturization and packaging. The most common form of control of SiPh POPs, thermal actuation, introduces unwanted side-effects of reciprocal thermal cross-talk between regions of the POPs, which, compounding with manufacturing tolerances, shifts the ideal expected working point of the POP, rendering the control of POPs a difficult task, and hindering their scalability and adoption. Therefore a complex control stack to compensate for these unwanted effects is required, and various control techniques have been presented, either using closed or open loop schemes, but artificial neural network-based approaches have not been thoroughly explored yet. This work proposes one such approach using an artificial neural network to solve the problem of control of thermally actuated POPs, by estimating the configuration of the phase-shifters as to realize a desired transfer function between input and output. This solution is applied on a SiPh POP based on Mach-Zehnder interferometers (MZIs) suffering from manufacturing tolerances and thermal cross-talk, when faced with beam-splitting, power-transfer tasks; the control stack is then tested on the same POP in an on-line manner, showing its actual real-world feasibility. The solution is then enhanced on simulated instances of ideal MZI-based POPs by introducing physics-informed constraints and embedding algorithms in the training of the artificial neural network, showing the improved performance of these techniques over the purely black-box direct model.
Sfruttando meccanismi di interferenza e modulazione, i processori ottici programmabili (POP) offrono una soluzione per la costruzione di dispositivi ottici in grado di svolgere operazioni direttamente nel dominio analogico. La loro implementazione hardware è favorita dall’utilizzo della infrastruttura di fabbricazione su silicio, permettendo, quindi, la loro miniaturizzazione e produzione, con costi contenuti. Il più comune meccanismo di configurazione di un POP in silicio, l’attuazione termica, introduce effetti indesiderati come il cross-talk termico tra regioni del circuito ottico, che, aggiungendosi a difetti di produzione, sposta il punto di funzionamento del sistema. Vari tipi di controllo sono stati presentati tramite meccanismi di retroazione, sia chiusa, sia aperta, ma soluzioni con reti neurali artificiali non sono ancora state esplorate e applicate per questo tipo di compito. Questa tesi propone un approccio al controllo dei POP attuati termicamente, tramite rete neurale artificiale, stimando i valori della configurazione del POP necessari per realizzare una funzione di transferimento desiderata. Questa soluzione è applicata a un caso reale di POP in silicio, costituito da una maglia di interferometri Mach-Zehnder, nella sua applicazione come combinatore lineare delle potenze in ingresso, verso l’uscita. Il sistema di controllo è, poi, verificato sul POP in maniera on-line, mostrandone la sua attuabilità. La soluzione è poi ampliata su istanze simulate del POP, introducendo meccanismi physics-informed e di embedding nell’addestramento della rete, presentando il loro miglioramento di prestazioni rispetto al puro approccio diretto.
Control of programmable optical processors through data-driven physically informed machine learning
Masini, Gabriele
2023/2024
Abstract
Exploiting light interference and modulation, programmable optical processors (POPs) allow the implementation of computations directly in the analog optical domain with on-premise, real-time reconfiguration of the behavior of the optical component. Through Silicon Photonics (SiPh), POPs take advantage of the well-consolidated manufacturing process used for electronic integrated circuits to fabricate optical components directly on silicon wafers, enabling cost-effective miniaturization and packaging. The most common form of control of SiPh POPs, thermal actuation, introduces unwanted side-effects of reciprocal thermal cross-talk between regions of the POPs, which, compounding with manufacturing tolerances, shifts the ideal expected working point of the POP, rendering the control of POPs a difficult task, and hindering their scalability and adoption. Therefore a complex control stack to compensate for these unwanted effects is required, and various control techniques have been presented, either using closed or open loop schemes, but artificial neural network-based approaches have not been thoroughly explored yet. This work proposes one such approach using an artificial neural network to solve the problem of control of thermally actuated POPs, by estimating the configuration of the phase-shifters as to realize a desired transfer function between input and output. This solution is applied on a SiPh POP based on Mach-Zehnder interferometers (MZIs) suffering from manufacturing tolerances and thermal cross-talk, when faced with beam-splitting, power-transfer tasks; the control stack is then tested on the same POP in an on-line manner, showing its actual real-world feasibility. The solution is then enhanced on simulated instances of ideal MZI-based POPs by introducing physics-informed constraints and embedding algorithms in the training of the artificial neural network, showing the improved performance of these techniques over the purely black-box direct model.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Masini_Master_Thesis.pdf
non accessibile
Descrizione: Master thesis
Dimensione
13.11 MB
Formato
Adobe PDF
|
13.11 MB | Adobe PDF | Visualizza/Apri |
2024_12_Masini_Executive_Summary.pdf
non accessibile
Descrizione: Executive summary
Dimensione
2.46 MB
Formato
Adobe PDF
|
2.46 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/10589/230070