An integrated photonic accelerator for 2D convolutions based on phase modulators

We demonstrate an integrated photonic two-dimensional convolution accelerator based on phase modulator (PM). By combining optical frequency combs, wavelength-division multiplexing (WDM), and delay line (DL) structures, we realize an efficient time–frequency-multiplexed convolution architecture. Compared with one-dimensional convolution architectures, the proposed scheme reduces redundant computations and significantly improves bandwidth utilization, achieving a threefold increase in computational speed under identical bandwidth conditions. Simulation results indicate that the accelerator enables high-precision image edge detection and attains 89.38% accuracy on the MNIST handwritten-digit classification task, thereby demonstrating its potential for practical applications.