Optoelectronic hybrid computing from handwritten digits to binary based on time-division multiplexed neural networks

Conventional silicon-based electronic computing is plagued by escalating power consumption and diminishing efficiency gains as device scaling approaches physical limits. Meanwhile, all-optical diffractive neural networks (DNNs) suffer from limited computational accuracy on specific tasks due to inherent physical noise and manufacturing imperfections. Optoelectronic neural networks (OENNs), which synergistically combine the ultra-low power consumption and massive parallelism of optics with the flexibility and robustness of electronics, have emerged as a promising paradigm to address these challenges. In this work, we propose a time-division multiplexed diffractive optoelectronic neural network (TDM-DOENN) for high-efficiency hexadecimal to binary calculation of handwritten digits. The architecture utilizes seven trainable time-sequential amplitude masks. These masks combine with the input digit images as time-sequential modulation. This design effectively decomposes the challenging 16-class classification task into four parallel 2-class classification channels. To further enhance system robustness against physical noise, we integrate a (7, 4) Hamming check mechanism with three additional check bits. We develop a two-stage training strategy with tailored activation function and pooling operation. This method achieves collaborative optimization between the optical and electronic components. It fully leverages their respective strengths and effectively suppresses speckle noise. Under theoretical conditions, TDM-DOENN achieves a 16-class classification accuracy of 91.97%, accompanied by a computational speed of 2.56 TOPs/s and an energy efficiency of 425.97 GOPs/W based on reliable numerical simulations. This work demonstrates significant potential for applications in high-speed visual sensing, optical communication systems, and low-power edge computing devices.