Simulating data in particle physics is expensive and not perfectly accurate. To get around this, researchers are now exploring the use of foundation models - large AI models trained in a general, task-agnostic way on large amounts of data. Just like how language models can be pretrained on the full dataset of Internet text before being fine-tuned for specific tasks, these models can learn from large datasets of particle jets, even without labels. After the pretraining, they can be fine-tuned to solve specific problems using much less data than traditional approaches.
In this study, researchers explored how one of the proposed foundation models originally trained on jets, can be repurposed for a very different challenge: reconstructing hadronically decaying tau leptons. This multi-task problem involves identifying sprays of particles that come from the tau decay from the background noise from other particles, reconstructing the kinematics of the decaying tau lepton and identifying its decay mode.
The results show that the pretrained model works well on this new type of data and outperforms models trained from scratch in several key aspects. This suggests that foundation models could help make particle physics analyses more efficient and scalable, especially in situations where high-quality simulated data is limited. The approach could be particularly useful for future studies involving rare particles like the tau lepton, including those in Higgs boson research.
10.21468/SciPostPhysCore.8.3.046
Computational simulation/modeling
Not applicable
Reconstructing hadronically decaying tau leptons with a jet foundation model
4-Jul-2025
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.