Extracting Signal Electron Trajectories in the COMET Phase-I Cylindrical Drift Chamber Using Deep Learning

TL;DR

This paper introduces a deep learning-based method for extracting signal electron trajectories in the COMET Phase-I cylindrical drift chamber, achieving high purity and retention rates amidst high background noise, and represents the first application of such techniques in this context.

Contribution

It presents the first application of deep learning for tracking in the COMET experiment, significantly improving signal extraction accuracy in a high-background environment.

Findings

Achieved 98% purity in signal cell identification.

Attained 90% retention rate for true signal hits.

Surpassed the design goal of 90% for both metrics.

Abstract

We present a pioneering approach to tracking analysis within the COMET Phase-I experiment, which aims to search for the charged lepton flavor violating $\mu\to e$ conversion process in a muonic atom, at J-PARC, Japan. This paper specifically introduces the extraction of signal electron trajectories in the COMET Phase-I cylindrical drift chamber (CDC) amidst a high background hit rate, with more than $40\,\%$ occupancy of the total CDC cells, utilizing deep learning techniques of semantic segmentation. Our model achieved remarkable results, with a purity rate of $98\,\%$ and a retention rate of $90\,\%$ for CDC cells with signal hits, surpassing the design-goal performance of $90\,\%$ for both metrics. This study marks the initial application of deep learning to COMET tracking, paving the way for more advanced techniques in future research.