Orbital angular momentum (OAM) offers a new degree of freedom for laser beams. The OAM beam has caught considerable attention in recent years due to its high-dimensional properties, demonstrating tremendous potential in cutting-edge fields such as super-capacity optical communication, rotational sensing, high-resolution imaging, optical information storage, and quantum technologies. The ability to diagnose OAM rapidly and precisely is crucial in these applications, involving OAM mode recognition and OAM spectral measurement. With the rapid development of artificial intelligence (AI) across various domains, leveraging AI technology has been considered a novel solution to OAM recognition. We review recent advances in OAM recognition based on AI technology from the perspective of AI model classification, with a focus on highlighting the research progress made by our team in this field. Additionally, we also discuss recent studies on AI-based OAM diagnosis under various disturbing scenarios.

Our review consists of three main sections. The first section provides a comprehensive overview of AI classifications, encompassing machine learning models, deep learning models, and hybrid learning models. It presents the fundamental characteristics of each category, providing relative information about the specific models of AI technology proposed in our study and numerical methods adopted in the hybrid learning models. Then, the basic OAM recognition principles are introduced, including the theory of identifying OAM modes within superposed OAM beams and the measurement of the constituent proportion coefficients for each mode, such as the OAM spectrum. This section serves as a foundational framework to provide readers with a thorough understanding of AI classifications and lay the groundwork for the subsequent review of OAM recognition.

In the second section, a systematic review of AI-based OAM recognition schemes is presented to discuss the schemes from the perspective of AI-based model classification. Previous studies are categorized based on the employed model types of machine learning models (Fig. 1), deep learning models (Fig. 2), and hybrid learning models (Fig. 3). We provide a comprehensive overview of previous approaches, analyze their strengths, and summarize the development trends. Additionally, we also concentrate on the contributions made by our team. Drawing inspiration from the powerful data processing capabilities of deep learning, we propose an adjusted ENN deep-learning model for OAM spectral measurement. A specially designed phase-only diffraction optical element is adopted to extract OAM features from the superposed OAM beam, and the neural network training is utilized to analyze the diffraction pattern to calculate the OAM spectrum. Under scenarios with seven superposed OAM modes, the OAM spectral measurement yields a root mean square (RMS) error as low as 10^-6. Furthermore, we propose a deep residual network (DRN)-based deep learning methodology to analyze the complex spectrum of a superposed OAM beam. The methodology can process up to 50 overlapping OAM modes within the range of [-150, 150], demonstrating exceptional performance with RMS errors reaching 0.002 for intensity spectra and 0.016 for phase spectra. Notably, the computational speed is significantly enhanced, reducing the processing time to mere 0.02 s. This remarkable improvement represents a nearly thousandfold increase in processing time compared to traditional helical harmonic expansion. Additionally, a scheme to directly emit multi-partite non-separable states from a laser cavity is proposed in another study, where we leverage a DRN to extract the phase shifting from interference patterns and thus measure the fidelity of classical non-separability. The groundbreaking research lays the foundation for related state tomography endeavors, underscoring that AI technology can validate classical non-separable characteristics among degrees of freedom like OAM.

The third section presents the recent advances in AI-based OAM recognition schemes under disturbances. It is recognized that the OAM beam can be distorted by disturbances during transmission, especially in non-uniform media such as the atmosphere and oceanic turbulences. However, in applications like optical communication and radar detection, it is essential to acquire the original emitted optical field information. Previous studies mitigate the influence of disturbances by introducing adaptive optics. Subsequent analysis and processing are then adopted to restore the original emitted optical field information. With the emergence of AI-based OAM recognition, leveraging AI technology to establish implicit representations between laser fields before and after disturbances holds the potential to provide a novel approach for obtaining the original optical information directly from distorted OAM beams. Additionally, we provide a summary of AI-based OAM modal sensing methods under disturbances, categorizing the discussions based on disturbance factors and utilization scenarios. Among these methodologies, our team demonstrated an AI-based distortion correction technique for vector vortex beams in 2020. The TACCNN network is designed and proposed for learning the mapping relationship between the intensity distribution of distorted vector vortex beams and the turbulent phase, which facilitates rapid and precise compensation. Notably, with a turbulence intensity parameter (D/r₀) of 5.28, this technique yields remarkable enhancement in mode purity which elevates from 19% to 70%.

Leveraging the powerful computational and learning capabilities of AI technology allows us to extract information from more complex OAM superposed modes, and accelerate data processing. The AI-based OAM recognition method stimulates breakthroughs in high-dimensional OAM control technology in fields such as communication and lasers. Despite initial success in detecting OAM modes, challenges persist in the rapid and high-precision computation and analysis of wide-mode-range OAM control, such as OAM combs and optical spatio-temporal vortices. The introduction of AI technology is expected to overcome the efficiency and accuracy limitations in traditional methods, as the complex phase structure design and extensive data analysis involved in high-dimensional OAM tailoring are aligned with the AI capability. Finally, we hope this review will provide valuable insights for people who are interested in AI-based OAM recognition and its applications, and inspire more novel and remarkable ideas.