Abstract: In recent years, the interdisciplinary deep integration of artificial intelligence (AI) technology with tropospheric duct research has been triggering a paradigm shift in the field of tropospheric radio wave propagation. This integration compels researchers to re-examine traditional modeling and sensing methods, fostering the vigorous development of two key research directions: “intelligent inversion prediction of tropospheric ducts” and “intelligent forecasting of tropospheric ducts”. Leveraging its powerful capabilities in nonlinear fitting and generalization, AI has gradually emerged as an efficient "intelligent detector and propagation simulator" for atmospheric ducts, providing adaptive environmental awareness and decision-making support for systems such as over-the-horizon communication, radar detection, and radio monitoring. Conversely, the unique channel propagation mechanisms of atmospheric ducts offer a natural testing ground for computational and sensing models based on the physics of electromagnetic waves, forming a beneficial complement to numerical computational methods. From a unified perspective, this article systematically reviews the fundamental concepts of radio wave propagation in tropospheric ducts, theoretical models, and research progress at the intersection of AI and tropospheric atmospheric ducts. In the aspect of “intelligent inversion prediction”, it focuses on the applications of AI technologies (particularly deep learning) in optimizing duct parameter inversion, achieving accurate duct type identification, constructing propagation prediction models, and revealing implicit relationships with meteorological elements. In the aspect of “intelligent forecasting”, the discussion centers on the application potential of AI in spatiotemporal refined modeling and multi-scale spatiotemporal prediction of ducts. Finally, the article summarizes current practical challenges, including data quality and scarcity, insufficient model generalization capability, and real-time deployment bottlenecks, and outlines promising frontier directions such as pan-domain dynamic sensing, end-to-end intelligent systems, deep learning generative models, and transfer learning.