Introduction
A graph neural network (GNN)-based method is developed for track finding in gamma-ray reconstruction within the HERD Fiber Tracker.
Materials and methods
Using Monte Carlo simulations at 0.5, 1 and 10 GeV gamma-ray, tracks are transformed into graph representations. Two edge labeling strategies are compared: The original labeling strategy treats all spatially connected edges from \(e^+e^-\) pair equally as signal, whereas the refined labeling strategy uses Monte Carlo information to reclassify non-physical edges intersecting between positron and electron trajectories as background, thereby suppressing “ghost edges” in pair production events. Systematic evaluation of four architectures—graph convolutional network, graph sample and aggregate (GraphSAGE), Multi-GraphSAGE and GraphSAGE with jumping knowledge (JK-GraphSAGE)—shows that JK-GraphSAGE has optimal edge classification performance by mitigating over-smoothing in deep networks through adaptive layer aggregation. Electron–positron pairs produced via photon conversion deposit tracks, which are reconstructed as distinct topological shapes. Topology aware linear fitting is applied to signal edges found by the GNN to reconstruct track endpoints. Although refined labels marginally reduce edge classification capability due to increased class imbalance, they enhance angular resolution by suppressing ghost edge influence.
Results
Under quality-filtered selection, the JK-GraphSAGE model achieves \(\theta _{68}=0.10^{\circ }\) at 10 GeV with a reconstruction efficiency of 97.8%.
Conclusion
These results validate the feasibility of GNN-based track finding, establishing a robust foundation for future end-to-end reconstruction frameworks that integrate track finding, fitting within unified deep learning architectures.
