Toward optically programmable crossbar arrays for array-level artificial neural network hardware

Neuromorphic computing has attracted considerable attention as an efficient approach to overcoming the limitations of conventional von Neumann architectures. For the practical hardware implementation of artificial neural networks (ANNs), device operation should extend beyond individual cells to array-level architectures, typically realized as crossbar arrays in which synaptic weights are stored as the conductance states of individual cells and directly utilized for matrix–vector multiplication (MVM). In this context, crossbar arrays serve not only as high-density integration platforms but also as core computing architectures in which stored conductance distributions are directly utilized for MVM. However, conventional electrically modulated arrays often suffer from unintended disturbances through shared interconnects during programming and erasing, thereby limiting computational reliability. As a promising alternative, optically modulated systems enable selective and parallel weight control in a non-contact manner. Realizing these advantages in practical computing, however, requires optical learning to be implemented at the array level. In this Perspective, we review optically programmable crossbar arrays from the perspective of array-level ANN implementation. We first discuss the key mechanisms governing optical programming and erasing, and then examine the major requirements for array-level optical learning, including update uniformity, reliability, and efficiency. Finally, we highlight current progress, remaining challenges, and future perspectives toward the realization of optically programmable crossbar array-based ANN hardware.