Abstract <p>Particle track reconstruction is a pivotal task in modern high-energy physics experiments. Traditional methods like the Kalman filter, though effective, face significant challenges in scalability and computational efficiency in environments with high track multiplicity and noise. To address these limitations, the TrackML competition was established to discover new, effective approaches for reconstructing particle trajectories with both high performance and quality. In this work, we introduce TrackNET, a deep learning model based on a GRU recurrent neural network architecture that reconstructs particle tracks by concurrently processing multiple seeds, beginning from hits on the first detector layer and iteratively predicting the region on the next detector layer where subsequent hits are likely to appear, thereby constructing the track. Unlike the Kalman filter, TrackNET avoids complex algebraic computations and has a minimal memory footprint, processing only a small subset of hits at a time. This efficiency enables the parallel execution of hundreds of models to generate a list of track candidates with high recall. These candidates are then ranked based on a relevance criterion to identify the most accurate tracks. When applied to the TrackML dataset, TrackNET achieved promising results in both processing speed and reconstruction accuracy. These results highlight TrackNET’s potential as a scalable, efficient solution for particle tracking, with promising implications for future collider experiments and high-energy physics research.</p>