Abstract
With data digitization becoming the new norm, demands for more efficient computing hardware are on the rise to cope with the exponential increase in data to process. Photonic data processors offer a promising solution as they can perform parallel data processing efficiently. However, Photonic Integrated Circuit (PIC) technologies, which are exploited to implement photonic components, require further development to implement data processing components. One key design criterion for computing hardware components is reconfigurability as this function allows data storage and processing to be implemented. This thesis hypothesizes that Chalcogenide Phase Change Materials (PCMs) can be used to tune PIC devices efficiently, thus developing efficient photonic processors. State-of-the-art tunable PIC technologies are volatile, which can be inefficient for data storage purposes. The unique bistable amorphous and crystalline structural states with distinguishable optical properties of PCMs allow non-volatile PIC devices to be developed. We envisage PCM-tuned devices to complement current tunable devices when developing the next generation of computing hardware. Hence this thesis aims to design, develop, and prototype PCM-tuned photonic switching devices to implement the computing hardware components.
A systematic two-step approach was used to develop the PCM-tuned devices. Design works were being carried out to optimize the PCM layer and device dimensions, followed by fabricating and operating the PCM-tuned devices. The choice of PCM used to tune PIC devices was determined by a comparison analysis done on four different PCM-tuned optical directional coupler devices: Ge2Sb2Te5, Ge2Sb2Se4Te1, Sb2S3, and Sb2Se3. Our results showed that Sb2S3 gave the best overall performance as it had the lowest optical loss amongst the four PCMs. However, we also identified challenges when incorporating them into PIC devices. The crystalline films consisted of anisotropic large birefringent crystal grains, which made it difficult to implement multiple switching states reliably. Moreover, Sb2S3 had the smallest change in refractive index upon structural phase transition, which led to large device footprints. We attempted to address these challenges when fabricating those devices.
Upon understanding the pros and cons of using the various PCMs to tune photonic devices, we went on to assess the feasibility of using PCM-tuned devices to implement hardware neural network (HNN) components used for image recognition applications. Non-volatile Sb2S3 directional coupler weighting devices were integrated with volatile Ge2Sb2Te5 microring resonator (MRR) non-linear activation function (NLAF) components to form a perceptron, which are basic building blocks of the HNN. A three-layer HNN model was designed using the PCM-based perceptron to recognize hand-written images from the Modified National Institute of Standards and Technology (MNIST) dataset. A training accuracy of 94.5% was achieved. Moreover, a two-temperature analysis on the devices showed that the perceptron operating speed was ≤ 1 ns. These results indicate that PCM-tuned devices show potential for implementing efficient computing hardware.
As we ascertained that PCM-tuned devices are feasible for implementing computing hardware, PCM-tuned device prototypes were subsequently fabricated. To focus more on the PCM, the device prototypes were developed from ready-made standard photonic devices obtained from a photonics foundry. The fabrication process was expedited as it bypassed the waveguide fabrication process. However, converting the standard devices to PCM-tuned devices was challenging as the thick top oxide cladding layer, used for waveguide mode confinement, had to be removed for PCM deposition. The oxide removal process must ensure that the waveguide surface is smooth to minimize losses. A fabrication process, which involved the oxide etching process, was ultimately established to modify the foundry chips into PCM-tuned devices.
The developed MRR devices were then used to study how the robustness of Sb2S3 and Ag-infused Sb2S3 can be improved in PIC devices. In particular, we hypothesized that caps could address the limitations of Sb2S3, which were identified during the device design process. Our experimental results with (ZnS)0.8-(SiO2)0.2 caps supported this hypothesis. A non-isothermal crystallization study showed that the crystal grain size in capped films was substantially reduced. Interestingly, we also found that the loading effect from the capping layer resulted in greater spatial mode confinement on the PCM layer. The larger spatial confinement in the PCM layer resulted in larger device tunability, which in turn reduced the overall device footprint.
Based on the above findings, we concluded that PCMs can be used to tune PIC devices and develop efficient computing hardware. We envisage this technology to play an influential role in the development of next-generation computing hardware.