Sparse weight mapping and computation reuse strategy for scalable photonic matrix multiplication

Based on phase-change materials (PCMs), the photonic crossbar array offers non-volatile reconfigurability and high integration density, enabling efficient large-scale parallel photonic matrix multiplication. However, its scalability is fundamentally limited by optical transmission loss, impeding the practical implementation of large-scale matrix multiplication. To overcome this limitation, this study proposes a weight mapping strategy that enables larger convolutional computation to be executed efficiently within a scale-limited photonic crossbar array. A high-quality 4 × 4 photonic crossbar array with 3-bit precision modulation has been fabricated. Applied to an edge detection task, the mapping strategy encoded four different 3 × 3 operators onto the 4 × 4 photonic crossbar array, achieving a 225% improvement in computational efficiency. Moreover, when integrated into a photonic convolutional neural network, the strategy delivered a 96.7% classification accuracy on the MNIST dataset, showing excellent agreement with the theoretical simulation result of 96.84%. Our work opens a path toward large-scale photonic matrix multiplication under hardware constraints, advancing the development of photonic computing.