Track Reconstruction Transformer: A Scalable Deep Learning Approach for Particle Tracking in High-Energy Physics

Particle track reconstruction is crucial for high-energy physics, particularly in high-luminosity environments where traditional methods encounter limitations in scalability and efficiency. To overcome these challenges, we introduce the Track Reconstruction Transformer (TRT), a novel deep learning model inspired by transformer models for the object detection that offers significant advancements in particle trajectory prediction. TRT leverages a Transformer architecture to process detector hits and directly predict particle track parameters, generating a set of potential tracks, each with an associated probability indicating the probability that it represents an actual track to account for the variable number of tracks per event. This approach adopts the parallel processing efficiency of DETR, offering significant speed and scalability advantages compared to conventional methods. The TRT model has been evaluated using a toy dataset with helical track approximations and uniform noise hits. We present details on the model’s implementation, training, and performance, including strategies to address the $$O({{N}^{2}})$$ complexity of transformers. While current tests focus on processed hits, we also discuss the potential for adapting TRT to raw detector data (avoiding the hit reconstruction phase) in future work, leveraging the flexibility of transformer architectures for even broader applicability.