"A multi-mode fiber supports thousands of transmission modes," Qin explained. "In the actual transmission process, there is coupling between the modes, which can be equivalent to a multiple scattering process."

Therefore, small differences between multiple signals at the input end will be amplified by this multiple scattering physical process, presenting as speckles with different characteristics at the output end.

These speckle differences are difficult for the human eye to recognize, and deep learning happens to be very good at identifying such features, so that non-orthogonal input signals can be demultiplexed.

The researchers take a fully connected layer similar to the multi-input multiple-output relationship of multimode fiber to series the convolutional layer with speckle recognition characteristics.

Combined with the physical characteristics of multiple scattering in multimode fiber, the signals with slight differences at the input end of multimode fiber are converted into speckle recognizable by the neural network at the output end.

Then, the network is trained by the natural image data set and its corresponding output speckle. The trained network can recover multiple non-orthogonal multiplexed input signals only through a single output speckle, and then the multiplexed transmission of non-orthogonal input signals is realized.

For a long time, Qin Yuwen's team has been engaged in the research of optical communication, optical sensing and synaesthesia fusion photon technology. Previously, researchers have accumulated a certain technical foundation in studying the use of multimode fiber as a transmission medium.

In 2022, it was reported that deep neural network was used as the decoder of multi-mode fiber orthogonal multiplexing transmission system, and it was found that the combination of fully connected neural network and U-type convolutional neural network could improve the decoding fidelity [2].

In 2023, the research group re-examined the deterministic physical method based on the transmission matrix, and used the transmission matrix to directly invert, and obtained nearly perfect amplitude-phase transmission fidelity in multi-mode fiber transmission [3].

In this study, although the experimental results show that the proposed network has good generalization ability, it still relies on the training of a large number of data and lacks generalization and interpretability of neural networks, which is not conducive to the practical application of deep neural networks in the field of non-orthogonal multiplexing transmission technology.

The team said: "We are currently working on integrating multiple scattering physics further into the structure of the neural network. We hope that with a deeper understanding of the physical mechanism, we can train the network with a small amount of data."

On the other hand, the transmission characteristics of multimode fiber are more sensitive to environmental parameters. Therefore, the desensitization of environmental parameters by non-orthogonal multiplexing transmission technology is also one of the technical problems that need to be optimized, and researchers plan to solve this problem by updating the network in real time [4].

"This study is an important step forward in the field of multimode fiber non-orthogonal multiplexing transmission, and there is still a lot of work to continue to explore in order to truly push the technology to the application stage." Qin Yuwen said.

Next, the team will study the coupling characteristics between multimode fiber modes in the case of disturbance, and further explore the physical process and analysis method of multimode fiber multiple scattering.

They hope that the physical properties of the multiple scattering system can guide the optimization direction of the deep learning network structure, and improve the accuracy, rate and stability of the multi-mode fiber non-orthogonal transmission system.

In addition, it is expected to extend the technology to other multiple-input multiple-output wave systems, thereby increasing the transmission capacity limit of a single transmission medium.National Science and Technology Zhou Tongliang Library launched children's science experience activities

Chongqing, May 25 (People's Daily) -- From May 25 to June 1, is the 2024 National Science and Technology Week. In order to allow children to discover science, understand science and explore science,

on the 25th, Chongqing Tongliang District Library and Chongqing Just Psychological Consulting Co., Ltd. held the "scientific wings flying childhood dreams" children's science experience activity, and simultaneously opened the "carrying forward the spirit of scientists to stimulate the innovation of the whole society" 2024 National Science and Technology Activity Week.

"Why are there seasons of spring, summer, autumn and winter? Why does boiling water give off white gas? ..." At the activity site of Tongliang District Library, just 30 children were led by psychological counselors to observe science in life through interesting experimental phenomena,

stimulate children's enthusiasm for learning, and let children understand the scientific principles behind the experiment; By exploring scientific principles, help children upgrade from playing science to researching science, guide children to learn to explore and set goals.

Pu Keling, director of Tongliang District Library, said that in order to strengthen the construction of regional science popularization capacity and deeply implement the national scientific quality action,

Tongliang District Library will hold a series of colorful and diverse activities during the annual science and technology activity week. The science and technology week includes six series of online and offline activities,

including children's science experience, science knowledge competition challenge, reading punch, science and technology activities into the school, fun science small classroom, etc., aiming to let children learn more science and technology knowledge and feel the extraordinary charm of science.