Learning-enabled recognition of LG beams from multimode fiber specklegrams

The use of orbital angular momentum (OAM) beams, like Laguerre Gaussian (LG) beams is a significant advancement in optical communications, enhancing versatility in data encoding. In optical communications, especially OAM multiplexing, previous studies have successfully recognized LG beams in free space. However, free space propagation presents challenges including power loss, beam distortion, wandering, and physical obstacles. To overcome these hurdles, we introduce a novel method using speckle patterns (specklegrams) within a multimode fiber (MMF) for LG beam recognition through deep learning. Our study focuses on recognizing 15 $L{G}_{p|l| }$ beams ($0\le p,|l|\le 3$; excluding $L{G}_{00}$) generated with a spatial light modulator (SLM) and launched into a standard silica MMF with a 200 µm core diameter, producing corresponding speckle patterns. We utilize the widely-used deep convolutional neural network (CNN) model, AlexNet, to train these specklegram images, achieving an impressive 91 % recognition accuracy. Furthermore, we assess the recognition accuracy of 15 $L{G}_{p(–l)}$ modes, obtaining an 86.5 % accuracy rate. We also explore the impact of external perturbations, such as transverse weights (ranging from 0 to 3 kg in 1 kg increments) and temperature variations (ranging from 50 °C to 150 °C in 10 °C steps) on a section of MMF. In both scenarios, recognition accuracies are > 80 %. These findings highlight the effectiveness of our proposed approach, promising high-capacity data transmission in optical communications.