Development of radiation-tolerant beam imaging via multimode fiber and synthetic data-driven machine learning

Transverse beam profile monitoring in high-radiation areas is challenging due to camera degradation. A proposed solution employs a multimode fiber (MMF) to relay optical signals from the radiation zone to a shielded area, where a standard CMOS camera can operate safely. However, MMF transmission introduces significant distortions, producing complex speckle patterns at the output. We present a machine-learning–based method to reconstruct the transverse beam distributions from these patterns. The model was trained solely on synthetic data generated via stochastic Gaussian mixture simulations, in which the samples were displayed on a laser-illuminated digital micromirror device (DMD) and relayed through a 5-m MMF. The same setup was used for testing, with beam images on a scintillating screen from CERN’s CLEAR facility replayed on the DMD. The model achieved a 1.49% mean relative root mean square error across four key beam parameters—77% lower than the 6.59% baseline, demonstrating its potential for diagnostics in radiation-constrained environments.