Integrated weighting and demultiplexing photonic circuits on SOI for optical CNNs

Convolutional neural networks (CNNs) underpin many modern image and video processing systems. As these models grow in depth and complexity, they increasingly strain the capabilities of state-of-the-art digital hardware. Optical computing offers a promising alternative, thanks to its broad bandwidth, low latency, and the absence of capacitive charging limitations. In particular, time–wavelength interleaving provides an efficient way to map convolutional operations onto intertwined temporal and spectral domains, enabling high throughput signal processing. For scalable implementations, however, individual wavelength channels still need to be modulated by integrated de-multiplexing and wavelength-selective weighting elements. Herein, we present device architectures for de-multiplexing and weighting individual wavelength channels, integrated on a silicon-on-insulator platform using the imec iSiPP50G foundry process, thereby enabling compact and scalable optical neural-network computation.