Deep photonic networks with arbitrary and broadband functionality

The increasing application space in optical communications, computing, and sensing drives the demand for high-performance integrated photonic components. Designing on-chip systems with complex and application-specific functionality goes beyond the limits of physical intuition alone. Consequently, machine learning-based design methods have gained popularity in recent times. However, the computational requirements for accurate device simulations are expensive, posing a critical challenge. As a result, these methods often have limitations in terms of scalability and the degrees of freedom they offer for optical design in application-specific and arbitrary photonic integrated circuits. To address these challenges, in this thesis, a highly scalable physics-informed framework for designing on-chip optical systems with arbitrary functionality is proposed. The framework is based on a deep photonic network comprising custom-designed Mach-Zehnder interferometers. By utilizing this framework, ultra-broadband power splitters and a spectral duplexer are successfully designed within a remarkably short span of fewer than two minutes. Furthermore, state-of-the-art experimental performance for all devices is demonstrated, achieving insertion loss of less than 0.66 dB and a 1-dB bandwidth exceeding 120 nm. The presented framework serves as an essential tool, offering a feasible approach towards systematically designing large-scale photonic systems. It enables the customization of power, phase, and dispersion profiles, thus catering to the diverse requirements of multi-band optical applications. Such applications include high-throughput communications, quantum information processing, and medical/biological sensing.