Optical end-to-end implementations for scalable photonic computing

Thesis

Integrated photonics is increasingly explored as a platform for high-speed and energy-efficient computing, yet its scalability remains constrained by the persistent reliance on electronic interfaces to perform the accumulation of multiplexed optical signals and to supply nonlinear activation, thereby limiting cascadability and system bandwidth. This thesis introduces a new class of photonic devices that repurpose heat—traditionally regarded as a parasitic and slow effect—into a functional computational mechanism. Central to this work is the Photonic-Heater-in-Lightpath (PHIL) architecture, in which subwavelength nanoheaters embedded within silicon microring resonators convert dissipated optical energy directly into controllable thermo-optic phase shifts.

The thesis establishes three key capabilities of this architecture. First, it presents the inaugural demonstration of cross-wavelength all-optical encoding and summation, showing that optical power carried at multiple control wavelengths can be selectively transduced onto a spectrally distinct probe wavelength. This establishes PHIL as a mechanism for wavelength-routed accumulation and transformation of optical signals. Second, by exploiting the intrinsic Lorentzian resonance profile of the microring, the device produces continuously reconfigurable nonlinear transfer characteristics. These analogue activation functions emulate neuronal behaviour and provide the nonlinear mapping required for neuromorphic photonic systems. Third, the slow thermal relaxation of the nanoheater is shown to operate as a physical analogue integrator for ultrafast optical signals. The device performs leaky temporal integration of high-repetition-rate pulse trains over a 250 ns thermal window, enabling temporal accumulation with concurrent nonlinear activation entirely within the optical domain.

Together, these results position opto-thermo-optic interactions as a foundational computational resource for integrated photonics. By unifying cross-wavelength encoding and summation, reconfigurable nonlinear activation, and analogue temporal integration within a single CMOS-compatible device, the PHIL platform establishes a coherent framework for all-optical processing across wavelength and time domains. The work demonstrates that slow thermal dynamics can be strategically engineered, not circumvented, to achieve functions difficult or impractical to realise using conventional high-speed modulators. The thesis concludes by outlining future opportunities enabled by this paradigm, including scalable wavelength-multiplexed neuromorphic processors, large-vector optical computing architectures, and fully optical computational pipelines built from densely integrated PHIL-based modules.

Authors: Zhang, Y

• DOI: 10.5287/ora-vydx8e6qg
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: integrated photonics; photonic computing; all-optical signal integration; programmable optical nonlinearities; neuromorphic computing
• Subjects: Photonic computing; Nanophotonics; Photonics