Orbital angular momentum multiplexing diffractive neural networks for high-capacity optical inference

Deep diffractive neural networks have emerged as an all-optical computing paradigm that overcomes the bottlenecks of traditional electronic computing. However, traditional deep diffractive neural networks are typically limited to single-task processing. Here we demonstrate an orbital angular momentum multiplexing diffractive neural network (OAM-MDNN) that completes several recognition tasks simultaneously by exploiting orthogonal OAM modes to encode and decode information. We develop an end-to-end auto-optimizing strategy that improves mode utilization efficiency by approximately 14 times and the peak signal-to-noise ratio by approximately 14.4% compared with manual encoding. We experimentally demonstrate a 10-mode multiplexing recognition task for 40 object categories, including handwritten digits, letters and fashion items, with an average recognition accuracy of 86.8%. Further experiments validate the scalability and performance of the network for up to 100 object categories. Theoretically, we also show that OAM-MDNN can be scaled to up to 40 modes times 10 classes (400 classifications), showing potential for further extension. In addition, our OAM-MDNN supports simultaneous multitask input, enabling parallel recognition with high accuracy and minimal intermode crosstalk. We applied our architecture to real-world tasks, such as gesture recognition in numerical simulations, which demonstrated its practical application potential in biometric identification. OAM-MDNN paves the way for high-capacity and multi-dimensional parallel optical information processing based on OAM multiplexing.