Graph Neural Network with Attention and Two-Stage Aggregation for Particle Track Reconstruction in the TPC MPD of the NICA Accelerator Complex

A Graph Neural Network (GNN) with an attention mechanism and two-stage aggregation for particle track reconstruction in the Time-Projection Chamber (TPC) of the Multi-Purpose Detector (MPD) of the Nuclotron-based Ion Collider fAcility (NICA) is presented. The model is trained and evaluated on a dataset of 1000 Au–Au collision events generated with MPDRoot. The model architecture includes node and edge encoders, an initial edge classifier, a graph attentional convolution layer, an edge embedding updater, and a final edge classifier, with the core blocks repeated for iterative refinement. A comparative analysis shows that the introduction of the attention mechanism and two-stage aggregation improves performance in edge classification, achieving 96.2% accuracy along with 92.6% in both purity and efficiency metrics. The track reconstruction efficiency exceeds 90% for track integrities below 80%, but rapidly drops for higher integrity thresholds.