Tau hadronic decay channel reconstruction with a Lorentz-Covariant Graph Attention Network

Abstract The reconstruction of hadronic decay modes of the τ lepton is a key task in high-energy physics experiments, directly impacting precision measurements and new physics searches. Conventional reconstruction approaches rely on handcrafted features and rule-based algorithms, which face challenges in handling complex decay topologies and detector effects. In this work, we propose LCGA-Net, a physics-informed deep learning framework for τ decay channel classification and four-momentum reconstruction. The method represents decay products as particle graphs and employs a multi-head attention mechanism to capture inter-particle correlations. To ensure physical consistency, Lorentz-invariant constraints are explicitly incorporated into the model architecture. A dedicated single-τ dataset is constructed from the simulation data of the Circular Electron Positron Collider (CEPC), covering six major hadronic decay modes. The proposed model achieves an accuracy of 89.2% on the validation set, matching or slightly exceeding baseline models.More importantly, full physics constraints substantially improve the physical plausibility of reconstructed four-momenta—particularly the energy–momentum relation—without degrading classification performance. These results demonstrate that integrating physics priors with with attention-based deep learning on particle clouds provides an effective approach for τ decay reconstruction, with potential applications in future high-energy physics analyses.