jet-universe/sophon-ak4

Model Card: SophonAK4

The SophonAK4 model is a realistic small-radius (R = 0.4) anti-k_T jet tagger developed for fast-simulation (Delphes) datasets under the JetClass-II configuration, designed to emulate the CMS detector conditions at the LHC.

Here, realistic indicates that the model achieves tagging performance comparable to state-of-the-art jet taggers used in the ATLAS and CMS experiments.

The model is constructed to cover a broad range of final states, including partons and leptons of various flavors and charges.

Model Details

SophonAK4 is trained using a multi-class classification approach based on di-X resonance processes, where the resonance X decays into multiple two-prong final states. Truth labelling is performed by associating reconstructed anti-k_T jets with partons or leptons originating from these two-prong decays.

A total of 23 jet labels are defined:

• Single-prong labels: $bb$, $\bar{b}\backslash bar\{b\}$, $cc$, $\bar{c}\backslash bar\{c\}$, $ss$, $\bar{s}\backslash bar\{s\}$, $dd$, $\bar{d}\backslash bar\{d\}$, $uu$, $\bar{u}\backslash bar\{u\}$, $gg$, $e^{-}e^-$, $e^{+}e^+$, ${\mu }^{-}\backslash mu^-$, ${\mu }^{+}\backslash mu^+$, ${\tau }_{\mathrm{h}}^{-}\backslash tau\_\{\backslash rm\; h\}^-$, and ${\tau }_{\mathrm{h}}^{+}\backslash tau\_\{\backslash rm\; h\}^+$. These correspond to cases where a single truth particle (either a parton or a lepton) is matched to the jet within ΔR(jet, particle) < 0.4, while the other particle from the same resonance decay is not matched to the jet.

• Two-prong labels: $b\bar{b}b\backslash bar\{b\}$, $c\bar{c}c\backslash bar\{c\}$, $s\bar{s}s\backslash bar\{s\}$, $d\bar{d}d\backslash bar\{d\}$, $u\bar{u}u\backslash bar\{u\}$, and $gggg$. These labels are assigned when both particles from the same resonance decay are matched within the same jet.

Uses

Integrating SophonAK4/Sophon Models

The SophonAK4 model, together with the Sophon model, provides a realistic benchmark for small- and large-R jet tagging on fast-simulation (Delphes) datasets, achieving performance comparable to state-of-the-art taggers used in the ATLAS and CMS experiments.

• For an example of integrating them in C++ workflows to analyze Delphes files, check [here]. (note: the SophonAK4 model will be supported since April 25')

• For an example of how to integrate these models into the Delphes processing workflow, refer to the following GitHub repository: https://github.com/jet-universe/delphes/tree/jet-models (note: will be available since May 25')

Evaluation

The performance of SophonAK4 is evaluated using the standard model $t\bar{t}t\backslash bar\{t\}$ events to enable direct comparison with performance benchmarks from ATLAS and CMS. Details are provided in the [Appendix B of the paper], and are summarized below.

For b- and c-tagging, genuine b, c, and light-flavor jets are selected via jet-parton matching as implemented in Delphes. Jets are required to satisfy p_T > 30 GeV and |η| < 2.5, consistent with CMS configurations.

The following b-tagging discriminant is constructed from SophonAK4's the raw output scores to evaluate b vs. light and b vs. c jet performance:

• $\text{discr (SophonAK4 }b\text{ tagging)}=g_b+g_{\bar{b}}+g_{b\bar{b}}\mathrm{.}\backslash text\{discr\; (SophonAK4\; \$b\$\; tagging)\}\; =\; g\_\{b\}\; +\; g\_\{\backslash bar\{b\}\}\; +\; g\_\{b\backslash bar\{b\}\}.$

The following c-tagging discriminants are defined for c vs. light and c vs. b jets, respectively.

• $\text{discr (SophonAK4 }c\text{ tagging)}=g_c+g_{\bar{c}}+g_{c\bar{c}},\backslash text\{discr\; (SophonAK4\; \$c\$\; tagging)\}\; =\; g\_\{c\}\; +\; g\_\{\backslash bar\{c\}\}\; +\; g\_\{c\backslash bar\{c\}\},$
• $\text{discr (SophonAK4 }c\text{ vs. }b\text{ tagging)}=\frac{g_c+g_{\bar{c}}+g_{c\bar{c}}}{g_c+g_{\bar{c}}+g_{c\bar{c}}+g_b+g_{\bar{b}}+g_{b\bar{b}}}\mathrm{.}\backslash text\{discr\; (SophonAK4\; \$c\$\; vs.\; \$b\$\; tagging)\}\; =\; \backslash frac\{g\_\{c\}\; +\; g\_\{\backslash bar\{c\}\}\; +\; g\_\{c\backslash bar\{c\}\}\}\{g\_\{c\}\; +\; g\_\{\backslash bar\{c\}\}\; +\; g\_\{c\backslash bar\{c\}\}\; +\; g\_\{b\}\; +\; g\_\{\backslash bar\{b\}\}\; +\; g\_\{b\backslash bar\{b\}\}\}.$

1. The ROC performance for b vs. light/c jets and c vs. light/b jets is shown below and can be compared to CMS benchmarks (Figs. 1 and 3 for the tt̅ process).

Conclusion

• The b vs. light jet performance is slightly below that of the widely-adopted DeepJet tagger in CMS.
• The b vs. c and c vs. light/b jet performances fall between DeepJet and ParticleNet taggers in CMS.
• Similar trends are found by comparing with ATLAS's widely-adopted DL1r tagger, see Appendix B of the paper.

2. Performance across different p_T and |η| regions is benchmarked below and can be compared with CMS benchmarks (Figs. 17, 19, 21, 23, 25, 27, 29, and 31).

Conclusion

• Tagging performance degrades in the low-p_T and high-|η| regions but reaches the plateau beyond the turn-on point, indicating that the SophonAK4 tagger exhibits realistic flavor-tagging behavior across kinematic regimes.

Citation

If you find the SophonAK4 model useful in your research, please cite our [paper] that introduces the model:

@article{Zhao:2025rci,
    author = "Zhao, Yuzhe and Li, Congqiao and Agapitos, Antonios and Fu, Dawei and Gao, Leyun and Mao, Yajun and Li, Qiang",
    title = "{Novel $|V_{cb}|$ extraction method via boosted $bc$-tagging with in-situ calibration}",
    eprint = "2503.00118",
    archivePrefix = "arXiv",
    primaryClass = "hep-ph",
    month = "2",
    year = "2025"
}