Speaker
Description
Denoising and track reconstruction in drift chambers fundamental to particle identification and momentum measurement at electron-positron colliders. While Transformer architectures have revolutionized many sequence-processing domains, their potential for track reconstruction in high-energy physics has not been fully explored. In this work, we introduce Transformer-based methods at two stages of the reconstruction chain. At the hit level, we employ an attention-based model trained via masked learning to distinguish signal hits from background noise, achieving robustness across varying track multiplicities. At the track level, a fine-tuned BERT model performs hit-to-track association. We also evaluate the impact of input representations by comparing physical variables in detector space with latent features derived from Graph Neural Networks. This study illustrates a practical integration of advanced deep learning techniques into high-energy physics tracking systems, providing a promising pathway to enhance reconstruction efficiency and accuracy.