Evaluating machine learning models for multimode-fiber-based transverse beam profile reconstruction

Transverse beam profile monitoring is essential for safe and efficient accelerator operation. In high-radiation environments such as beam dumps, cameras degrade rapidly. To address this, a single multimode fiber (MMF) transmission system was previously tested to transport scintillation light from a screen to a remote camera. Because multiple guided modes are excited and coupled during propagation, the fiber output does not preserve the image and requires reconstruction. This contribution evaluates seven machine-learning reconstruction models for recovering the original transverse beam distribution from MMF output. Using data from the MMF-relayed Chromox screen campaign at CERN’s CLEAR facility, the study compares models in terms of reconstruction error, convergence speed, and run-to-run stability, with particular attention to the use of incoherent light. The results indicate robust options for radiation-tolerant, MMF-based transverse diagnostics.