Abstract
The rapid growth of artificial intelligence and large-scale data processing has exposed fundamental limitations in conventional electronic computing. Optical computing has emerged as a promising platofrm for information processing, offering inherent advantages in parallelism, power efficiency, and low-latency. Photonic metasurfaces, composed of subwavelength arrays of nanostructures, emerged as a compact and viable platform for this purpose, capable of modulating light with tailor-made amplitude, phase, and polarization. This thesis presents a systematic exploration of metasurface-based optical computing, advancing the platform from individual meta-atom design toward programmable, multifunctional optical processors for both classical and quantum information processing.
This begins by establishing the physical and numerical foundations of metasurface design, including Jones calculus, meta-atom designs, numerical tools, and fabrication techniques. Building upon these foundations, I show a metalens array platform is introduced for performing complex-valued linear algebra in the optical domain. A single-layer metalens array is experimentally demonstrated to execute the discrete Fourier transform (DFT) with interferometric phase retrieval and a pixel-level error-mitigation scheme. The same architecture is extended to matrix-matrix multiplication, implementing a two-pass optical JPEG compression algorithm based on the two-dimensional discrete cosine transform. Ability to select different sets of inputs and outputs channels is further exploited to simulate quantum algorithms, including Grover’s search and the quantum Fourier transform, with the results validated using both classical and heralded single-photon sources.
To address the limitations of forward-design and single-layer designs, specifically interference artifacts and sensitivity to pixel-level alignment, the thesis introduces a multi-layer holographic metasurface design via stochastic gradient descent (SGD) for complex holographic field engineering. This starts by a simple demonstration of generating a complex-valued holograms, a task conventionally constrained by the limited phase-only modulation of single interfaces. Then, the thesis extend this methodology to dual-layers, which effectively suppresses unwanted diffraction and eliminates the vortex-like artifacts inherent in discrete metalens arrays, resulting in a more robust platform for complex-valued matrix-vector multiplication (MVM). Furthermore, the thesis explores the function of metasurfaces into the quantum regime, to demonstrate the generation and reconstruction of Bell states holograms.
In the last result chapter, the thesis shift the interest from designing the metasurface to optimizing the device and scaling the methodology towards a larger scale computation. In this regard, a graph-theoretical optimization framework is then developed to minimize the physical footprint of metalens arrays required for simulating open quantum systems. Finally, the thesis addresses the nonlinearity gap in optical computing by proposing intensity-based projection layers as effective nonlinear activations. This framework enables the construction of complex-valued and quaternion-valued optical neural networks capable of image classification, reconstruction, and parameter extraction with accuracy comparable to conventional real-valued networks.
Collectively, this thesis establishes metasurfaces not merely as flat optical components, but as a foundational device for the next generation of hybrid optoelectronic computing. The findings and methodoloty presented here provide a scalable path toward future photonic processor capable of meeting the computational and energy demands of the future.