Monolithic Amplitude‐Phase Modulator for Scalable Optical Convolution

ABSTRACT Photonic neural networks (PNNs) are promising for next‐generation AI accelerators due to their inherent parallelism and ultra‐fast optical matrix‐vector multiplication capabilities. However, scaling these systems is fundamentally limited by the quadratic ( N 2 ) growth in modulator count required for N × N weight matrices, leading to prohibitive integration density and optical losses in large‐scale systems. We overcome this limitation through a novel hybrid MZI‐PIN‐PCM architecture that synergistically integrates Mach‐Zehnder interferometers (MZIs) with low‐loss phase‐change materials (PCMs) and waveguide‐integrated silicon p‐i‐n (PIN) microheaters. Our design simultaneously enables independent, high‐performance amplitude (>7‐bit precision, 100 MHz) and phase (non‐volatile, >600 cycles) modulation while reducing the required modulator count by 50% and achieving >10 dB lower optical losses for large matrices ( N >200) compared to conventional approaches. To demonstrate its functionality, we first validate on‐chip 2D vector multiplication (∼9‐bit precision) and then construct an optical convolutional neural network (CNN) using a four‐channel wavelength‐division multiplexing (WDM) system, with our modulator as the convolutional core. The system achieves recognition accuracies of 94.8% (MNIST) and 88.8% (Fashion‐MNIST), closely matching their electronic counterparts (97.0% and 89.8%, respectively). This hybrid architecture significantly advances PNNs by enhancing integration density, computational efficiency, and insertion loss reduction.