Space-division multiplexing using multimode fibers has emerged as a promising technique offering significantly enhanced data throughput on the modern internet. However, the use of multimode fibers introduces modal crosstalk due to complex transmission behavior. Accessing the exact modal weights via mode decomposition enables a full understanding of fiber transmission behavior and facilitates precise control of light propagation, enabling high-dimensional information transmission over optical fibers without distortion. This thesis presents a comprehensive investigation of AI–based approaches for the characterization of the optical field in multimode fibers using pure intensity measurement. By integrating AI algorithms with photonics, three strategies are proposed to achieve accurate, scalable, and real-time mode decomposition. The data-driven approach incorporates physical constraints into the training data, enabling accurate mode decomposition up to 910 Hz on 23 modes. Physics-informed models integrate physical priors directly into the optimization process, allowing decomposition of up to over one thousand modes. These numerical models demonstrate strong generalizability, operating effectively over fiber lengths ranging from 1 m to 1 km. For physics-in-network models, optical diffractive neural networks are digitally trained and physically implemented, enabling mode decomposition at the speed of light. These achievements in reference-less characterization and real-time demultiplexing of multimode fibers are expected to boost optical communication and high-dimensional quantum communication. Furthermore, these techniques are anticipated to broaden their impact across a range of applications, including integrated photonic devices, fiber sensing, and optical computing.