Neuro-explicit semantic segmentation of the diffusion cloud chamber

Abstract

For decades, in diffusion cloud chambers, different types of subatomic particle tracks from radioactive sources or cosmic radiation had to be identified with the naked eye which limited the amount of data that could be processed. In order to allow these classical particle detectors to enter the digital era, we successfully developed a neuro-explicit artificial intelligence model that, given an image from the cloud chamber, automatically annotates most of the particle tracks visible in the image according to the type of particle or process that created it. To achieve this goal, we combined the attention U-Net neural network architecture with methods that model the shape of the detected particle tracks. Our experiments show that the model effectively detects particle tracks and that the neuro-explicit approach decreases the misclassification rate of rare particles by 73% compared with solely using the attention U-Net.