Reconstruction of Unstable Heavy Particles Using Deep Symmetry-Preserving Attention Networks

Overview

• This paper presents an extended version of the Symmetry Preserving Attention Network (SPA-NET) architecture, which is used to reconstruct unstable heavy particles from detector data at the Large Hadron Collider (LHC).
• The extended SPA-NET model can handle multiple input object types, such as leptons and global event features, in addition to hadronic jets.
• The model provides both regression and classification outputs to improve the parton assignment process.
• The performance of the extended SPA-NET is evaluated on semi-leptonic top quark pair decays and top quark pairs produced in association with a Higgs boson.

Plain English Explanation

At the Large Hadron Collider, physicists study the collisions of particles to learn about the fundamental building blocks of the universe. When certain heavy, unstable particles are produced in these collisions, they quickly decay into other particles that are detected by the experiment. Reconstructing the original heavy particles from the detected particles is a challenging task due to the large number of possible configurations.

The researchers in this paper have extended a machine learning model called the Symmetry Preserving Attention Network (SPA-NET) to better handle this reconstruction problem. The extended SPA-NET model can take into account different types of detected particles, such as leptons and global event features like missing energy, in addition to the more common hadronic jets. The model also provides both classification and regression outputs to help assign the detected particles to the original heavy particles.

The researchers tested the extended SPA-NET model on two specific physics processes: the decay of top quark pairs and the production of top quark pairs along with a Higgs boson. They found that the extended model significantly improved the performance of three representative studies: a search for the Higgs boson produced with top quarks, a measurement of the top quark mass, and a search for a hypothetical heavy Z' particle decaying into top quark pairs.

Technical Explanation

The Symmetry Preserving Attention Network (SPA-NET) is a machine learning architecture that has been previously applied to the problem of reconstructing top quark pair decays at the LHC, which produce only hadronic jets. In this work, the researchers extend the SPA-NET architecture to handle multiple input object types, such as leptons, as well as global event features, such as missing transverse momentum.

The extended SPA-NET model provides both regression and classification outputs to improve the parton assignment process. The regression outputs predict the four-momenta of the underlying partons, while the classification outputs determine the most likely assignment of detector objects to partons.

The researchers evaluate the performance of the extended SPA-NET in the context of semi-leptonic top quark pair decays and top quark pairs produced in association with a Higgs boson. They find significant improvements in the power of three representative studies: a search for ttH, a measurement of the top quark mass, and a search for a heavy Z' decaying to top quark pairs.

The researchers also present ablation studies to provide insight into what the network has learned in each case. These studies can help identify the key features and mechanisms that contribute to the improved performance of the extended SPA-NET model.

Critical Analysis

The paper presents a well-designed extension of the SPA-NET architecture to handle more complex particle physics processes at the LHC. The inclusion of leptons and global event features, as well as the addition of regression and classification outputs, is a clear advancement over the previous version of the model.

However, the paper does not address certain limitations of the approach. For example, the model still relies on a predefined set of possible parton assignments, which may not capture all the possible configurations in real-world data. Additionally, the paper does not discuss the computational efficiency of the extended SPA-NET model, which is an important consideration for real-time applications at the LHC.

Furthermore, the paper could have provided more insights into the specific features and mechanisms that the network has learned, beyond the high-level ablation studies. A deeper understanding of the network's inner workings could lead to further improvements or inspire the development of alternative approaches.

Overall, the paper presents a promising extension of the SPA-NET architecture, but there are opportunities for further research to address the limitations and provide a more comprehensive understanding of the model's performance.

Conclusion

This paper describes an extended version of the Symmetry Preserving Attention Network (SPA-NET) architecture, which is used to reconstruct unstable heavy particles from detector data at the Large Hadron Collider. The extended SPA-NET model can handle multiple input object types, such as leptons and global event features, and provides both regression and classification outputs to improve the parton assignment process.

The researchers found that the extended SPA-NET model significantly improved the performance of three representative studies: a search for the Higgs boson produced with top quarks, a measurement of the top quark mass, and a search for a hypothetical heavy Z' particle decaying into top quark pairs. While the paper presents a promising advancement, there are opportunities for further research to address the model's limitations and provide a deeper understanding of its inner workings.