Abstract
Deep learning can give a significant impact on physics performance of electron-positron Higgs factories such as ILC and FCCee. We are working on two topics on event reconstruction to apply deep learning. The first is jet flavor tagging, in which we apply particle transformer to ILD full simulation to obtain jet flavor, including strange tagging. The second is particle flow, which clusters calorimeter hits and assigns tracks to them to improve jet energy resolution. We modified the algorithm developed in context of CMS HGCAL based on GravNet and Object Condensation techniques and add a track-cluster assignment function into the network. The overview and performance of these algorithms are described.
Introduction
Future electron-positron "Higgs factories," such as the International Linear Collider (ILC) and the Future Circular Collider (FCCee), demand unprecedented precision in event reconstruction to fully explore Higgs boson physics. Two critical areas for improvement are jet flavor tagging—specifically the challenging task of strange tagging—and the Particle Flow Algorithm (PFA). Traditional methods like LCFIPlus for tagging and PandoraPFA for clustering are reaching their limits in high-granularity detector environments. This work explores the application of state-of-the-art deep learning architectures to these challenges using the International Large Detector (ILD) full simulation.
Methodology
Jet Flavor Tagging with Particle Transformers
To improve jet flavor identification, we employ the Particle Transformer (ParT) architecture. Unlike standard Transformers used in natural language processing, ParT is designed with a "physics bias" that incorporates the spatial and kinematic relationships between particles. It processes both individual particle properties and an "interaction" term derived from the 4-momentum of particle pairs. This interaction term weights the attention mechanism, allowing the model to learn complex correlations within a jet rather than just treating particles as an unordered set.
Particle Flow via GravNet and Object Condensation
For Particle Flow, we utilize a distance-based Graph Neural Network (GNN) known as GravNet, combined with an "Object Condensation" loss function. This approach treats calorimeter hits and tracks as nodes in a graph. The network learns a latent coordinate space where hits belonging to the same particle are "pulled" toward a central condensation point. This allows for dynamic, end-to-end clustering and track-cluster association, bypassing the rigid step-by-step logic of traditional algorithms.
Experimental Setup
The study utilizes high-fidelity ILD full simulations generated with iLCSoft and Geant4, ensuring realistic modeling of detector effects and secondary interactions. Benchmarking was also performed against FCCee Delphes fast simulations. The models were trained on $e^+e^-$ collisions producing $ZH$ and $q\bar{q}$ events at center-of-mass energies ranging from 91 to 250 GeV. A key addition for strange tagging was the integration of "Comprehensive PID" (CPID), utilizing $dE/dx$ (energy loss) and Time-of-Flight data to distinguish high-momentum Kaons from Pions.
Key Results
The deep learning approach yielded significant performance improvements over traditional baselines:
- Massive Rejection Gains: In 3-category flavor tagging ($b, c,$ others), the Particle Transformer achieved a 5-10x improvement in background rejection compared to the previous state-of-the-art (LCFIPlus).
- Strange Tagging Breakthrough: The model demonstrated non-negligible statistical power in separating strange jets, a task previously considered extremely difficult without advanced PID integration.
- Superior PFA Efficiency: The GravNet-based PFA showed significantly higher efficiency and purity in jet environments compared to PandoraPFA. For example, in jet events, the GravNet approach reached an electron efficiency of 94.5%, compared to 80.2% for the traditional method.
Conclusion and Future Work
The application of Particle Transformers and GNNs to ILD simulations demonstrates the transformative potential of deep learning for Higgs factory physics. While clustering performance is high, future work will focus on improving energy regression—accurately determining the total energy of those clusters—which remains a challenge. Additionally, the team is exploring encoder-decoder Transformer frameworks to further enhance the flexibility and accuracy of particle flow reconstruction across diverse detector geometries.
Summary updated