Wavelength-multiplexed computation with a large-scale 16×16 silicon photonic processor

As electronic processors approach their limits in efficiency and scalability, integrated photonic processors offer a promising path for accelerating modern AI workloads. We present a 16×16 silicon photonic processor based on the Clements mesh—currently the largest experimentally demonstrated MZI-based system—capable of all-optical MNIST classification with class labels mapped to output ports. To overcome fabrication errors and crosstalk, we introduce a machine-learning-based calibration algorithm, achieving 79.1% test accuracy under compressed data. We further enhance throughput via wavelength-division multiplexing (WDM) using 95 wavelengths, demonstrating over 70% accuracy across 80 channels. A digital-twin model with a linear interaction matrix captures inter-phase-shifter crosstalk with 0.935 fidelity on random data, enabling robust generalization. Our results demonstrate a scalable and energy-efficient architecture for photonic inference, combining accurate learning, WDM parallelism, and predictive modeling—paving the way for general-purpose, wavelength-multiplexed optical AI accelerators.