Learning to increase matching efficiency in identifying additional b-jets in the $$\text {t}\bar{\text {t}}\text {b}\bar{\text {b}}$$ process

Cheongjae Jang, Sang-Kyun Ko, J. Choi, J. Lim, Yung‐Kyun Noh, T. J. Kim

Abstract

The $t\bar{t}H(b\bar{b})$ process is a critical channel for revealing Higgs boson properties, but it is hindered by a significant irreducible background from the $t\bar{t}b\bar{b}$ process. To improve the sensitivity of Higgs searches, it is essential to accurately distinguish between b-jets originating from top quark decays and "additional" b-jets from gluon splitting. This paper presents a deep learning approach that optimizes "matching efficiency"—the accuracy of identifying these additional b-jets—using specialized loss functions. Our method treats the entire event as a single unit, outperforming traditional binary classification by increasing matching efficiency from 62.1% to 64.5% in leading-order simulations.

The Challenge of $t\bar{t}b\bar{b}$ Background

In high-energy physics, studying the Higgs boson often involves its decay into a pair of bottom quarks (b-quarks) in association with a top quark pair ($t\bar{t}H(b\bar{b})$). This process is difficult to isolate because the $t\bar{t}b\bar{b}$ background process produces a nearly identical final state with at least four b-jets.

Feynman diagram of the $t\bar{t}b\bar{b}$ process, illustrating the production of top quark pairs and additional b-jets from gluon splitting.

As shown in the Feynman diagram above, identifying the "additional" b-jets is a complex task. Previous methods utilized binary classifiers to evaluate jet pairs independently. However, this approach ignores the global event structure, where exactly one "correct" pair must be identified among several candidates. This lack of context often leads to suboptimal matching performance.

Methodology: Event-Level Matching

We propose a shift from independent pair classification to event-level matching. By utilizing a permutation-equivariant Deep Neural Network (DNN), we can process all b-jet pairs in an event simultaneously. This allows the model to learn from the relative differences between all jets in the event.

To train these networks, we introduced four surrogate loss functions ($L_1$ through $L_4$) designed to be differentiable proxies for matching efficiency. For instance, $L_3$ employs categorical cross-entropy across all pairs in an event, encouraging the model to assign the highest score to the true additional pair. This "global" optimization ensures the model picks the best candidate within the event context rather than simply passing a fixed threshold for every candidate.

Results and Performance

We verified our approach using both synthetic toy data and realistic simulated $t\bar{t}b\bar{b}$ events from proton-proton collisions at $\sqrt{s} = 13$ TeV. The results demonstrate that direct optimization of matching efficiency consistently yields better results than standard binary cross-entropy ($L_{BCE}$).

Matching efficiency vs. training iteration for various loss functions. The proposed event-level loss functions ($L_1-L_4$) significantly outperform the binary classification baseline ($L_{BCE}$).

In Leading Order (LO) simulations, our method improved matching efficiency from 62.1% to 64.5%. In Next-to-Leading Order (NLO) simulations, which are more complex and realistic, efficiency improved from 59.9% to 61.7%. Beyond simple accuracy, the reconstructed physical variables (such as the invariant mass of the jet pairs) using our method showed significantly better agreement with theoretical truth distributions compared to the binary approach.

Conclusion

By shifting the focus from binary classification to event-level matching, we have achieved a measurable increase in the accuracy of identifying additional b-jets in $t\bar{t}b\bar{b}$ events. This methodology provides a more robust framework for background estimation in Higgs boson studies and will be a valuable tool for analyzing data from the LHC's upcoming high-luminosity phases.

Summary updated

Abstract

Abstract The $$\text {t}\bar{\text {t}}\text {H}(\text {b}\bar{\text {b}})$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mtext>t</mml:mtext><mml:mover><mml:mrow><mml:mtext>t</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mtext>H</mml:mtext><mml:mo>(</mml:mo><mml:mtext>b</mml:mtext><mml:mover><mml:mrow><mml:mtext>b</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mo>)</mml:mo></mml:mrow></mml:math> process is an essential channel in revealing the Higgs boson properties; however, its final state has an irreducible background from the $$\text {t}\bar{\text {t}}\text {b}\bar{\text {b}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mtext>t</mml:mtext><mml:mover><mml:mrow><mml:mtext>t</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mtext>b</mml:mtext><mml:mover><mml:mrow><mml:mtext>b</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:math> process, which produces a top quark pair in association with a b quark pair. Therefore, understanding the $$\text {t}\bar{\text {t}}\text {b}\bar{\text {b}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mtext>t</mml:mtext><mml:mover><mml:mrow><mml:mtext>t</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mtext>b</mml:mtext><mml:mover><mml:mrow><mml:mtext>b</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:math> process is crucial for improving the sensitivity of a search for the $$\text {t}\bar{\text {t}}\text {H}(\text {b}\bar{\text {b}})$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mtext>t</mml:mtext><mml:mover><mml:mrow><mml:mtext>t</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mtext>H</mml:mtext><mml:mo>(</mml:mo><mml:mtext>b</mml:mtext><mml:mover><mml:mrow><mml:mtext>b</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mo>)</mml:mo></mml:mrow></mml:math> process. To this end, when measuring the differential cross section of the $$\text {t}\bar{\text {t}}\text {b}\bar{\text {b}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mtext>t</mml:mtext><mml:mover><mml:mrow><mml:mtext>t</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mtext>b</mml:mtext><mml:mover><mml:mrow><mml:mtext>b</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:math> process, we need to distinguish the b-jets originating from top quark decays and additional b-jets originating from gluon splitting. In this paper, we train deep neural networks that identify the additional b-jets in the $${\text {t}}{\bar{\text {t}}}{\text {b}}{\bar{\text {b}}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mtext>t</mml:mtext><mml:mover><mml:mrow><mml:mtext>t</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mtext>b</mml:mtext><mml:mover><mml:mrow><mml:mtext>b</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:math> events under the supervision of a simulated $$\text{t}\bar{\text{t}}\text{b}\bar{\text{b}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mtext>t</mml:mtext><mml:mover><mml:mrow><mml:mtext>t</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mtext>b</mml:mtext><mml:mover><mml:mrow><mml:mtext>b</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:math> event data set in which true additional b-jets are indicated. By exploiting the special structure of the $$\text {t}\bar{\text {t}}\text {b}\bar{\text {b}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mtext>t</mml:mtext><mml:mover><mml:mrow><mml:mtext>t</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mtext>b</mml:mtext><mml:mover><mml:mrow><mml:mtext>b</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:math> event data, several loss functions are proposed and minimized to directly increase matching efficiency, i.e., the accuracy of identifying additional b-jets. We show that, via a proof-of-concept experiment using synthetic data, our method can be more advantageous for improving matching efficiency than the deep learning-based binary classification approach presented in [1]. Based on simulated $$\text {t}\bar{\text {t}}\text {b}\bar{\text {b}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mtext>t</mml:mtext><mml:mover><mml:mrow><mml:mtext>t</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover><mml:mtext>b</mml:mtext><mml:mover><mml:mrow><mml:mtext>b</mml:mtext></mml:mrow><mml:mrow><mml:mo>¯</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:math> event data in the lepton+jets channel from pp collision at $$\sqrt{s}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msqrt><mml:mi>s</mml:mi></mml:msqrt></mml:math> = 13 TeV, we then verify that our method can identify additional b-jets more accurately: compared with the approach in [1], the matching efficiency improves from 62.1 $$\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mo>%</mml:mo></mml:math> to 64.5 $$\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mo>%</mml:mo></mml:math> and from 59.9 $$\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mo>%</mml:mo></mml:math> to 61.7 $$\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mo>%</mml:mo></mml:math> for the leading order and the next-to-leading order simulations, respectively.
BibTeX
@article{jang2022learning,
  author = {Cheongjae Jang and Sang-Kyun Ko and J. Choi and J. Lim and Yung‐Kyun Noh and T. J. Kim},
  title = {{Learning to increase matching efficiency in identifying additional b-jets in the $$\textbackslash\{\}text \{t\}\textbackslash\{\}bar\{\textbackslash\{\}text \{t\}\}\textbackslash\{\}text \{b\}\textbackslash\{\}bar\{\textbackslash\{\}text \{b\}\}$$ process}},
  journal = {The European Physical Journal Plus},
  year = {2022},
  doi = {10.1140/epjp/s13360-022-03024-8},
}
Figure