Scaling up for end-to-end on-chip photonic neural network inference.

Optical neural networks are emerging as a competitive alternative to their electronic counterparts, offering distinct advantages in bandwidth and energy efficiency. Despite these benefits, scaling up on-chip optical neural networks for end-to-end inference is facing significant challenges. First, network depth is constrained by the weak cascadability of optical nonlinear activation functions. Second, the input size is constrained by the scale of the optical matrix. Herein, we propose a scaling up strategy called partially coherent deep optical neural networks (PDONNs). By leveraging an on-chip nonlinear activation function based on opto-electro-opto conversion, PDONN enables network depth expansion with positive net gain. Additionally, convolutional layers achieve rapid dimensionality reduction, thereby allowing for an increase in the accommodated input size. The use of a partially coherent optical source significantly reduces reliance on narrow-linewidth laser diodes and coherent detection. Owing to their broader spectral characteristics and simpler implementation, such sources are more accessible and compatible with scalable integration. Benefiting from these innovations, we designed and fabricated a monolithically integrated optical neural network with the largest input size and the deepest network depth, comprising an input layer with a size of 64, two convolutional layers, and two fully connected layers. We successfully demonstrate end-to-end two-class classification of fashion images and four-class classification of handwritten digits with accuracies of 96% and 94%, respectively, using an in-situ training method. Notably, performance is well maintained with partially coherent illumination. This proposed architecture represents a critical step toward realizing energy-efficient, scalable, and widely accessible optical computing.