On‐Chip Photonic Neural Network Architectures

ABSTRACT

Artificial intelligence is rapidly advancing, driving an urgent demand for computing hardware that can overcome the limitations of conventional electronic architectures. Photonic neural networks emerge as a promising alternative by leveraging the intrinsic advantages of light, including high bandwidth, parallelism, and low latency. This review provides a comprehensive and critical overview of on‐chip photonic neural networks, spanning fundamental device technologies, system architectures, and emerging applications. We examine key photonic building blocks, including Mach–Zehnder interferometer meshes, microring resonator arrays, metastructures, and phase‐change material cells, and analyze their respective strengths, limitations, and scalability challenges. Building on these platforms, we compare major neural network implementations, such as fully connected, convolutional, recurrent, reservoir, Transformer‐based, and spiking architectures, highlighting trade‐offs in precision, energy efficiency, and integration complexity. We further discuss application domains ranging from optical signal processing to quantum computing. Finally, we identify critical bottlenecks, including scalability, nonlinearity implementation, and large‐scale integration, and outline future directions toward practical, high‐performance photonic neural network systems. This work establishes a unified perspective that connects device‐level innovations with system‐level functionality, providing guidance for the development of next‐generation photonic computing technologies.