Deep learning-based restoration of noise-corrupted and saturated beam profiles for real-time proton beam monitoring and quality assurance

Abstract

The Korea Multi-purpose Accelerator Complex (KOMAC) operates a 100 MeV proton linear accelerator, providing a high flux proton beam at the TR103 general-purpose irradiation facility. A real-time and in situ proton beam profile monitoring system including a P43 phosphor screen and CMOS camera has been installed and tested at the facility. However, to ensure beam profile quality, two types of degradation must be addressed: beam profile noise and saturation. High background noise, combined with pepper noise from secondary radiation exposure, results in a noise-corrupted beam profile, while high flux proton irradiation causes beam profile saturation, truncating the upper portion. To effectively restore noise-corrupted and saturated beam profiles to their true beam profiles, we propose a deep learning-based beam profile restoration method that employs virtual beam profile datasets, with which large amounts of data can be acquired efficiently to increase model accuracy. We optimized deep learning models based on U-Net and ResNet architectures and evaluated the model performance applying the proposed method to restore both noise-corrupted and saturated beam profiles.