Particle-level transformers for 95 GeV Higgs boson searches at future $e^+e^-$ Higgs factories

Motivated by several mild excesses around 95~GeV, we investigate the prospects for a light scalar $S$ produced via Higgsstrahlung, $e^+e^- \to Z(\mu^+\mu^-)S$, at future $e^+e^-$ Higgs factories. We take the CEPC as a benchmark, with a center-of-mass energy of $\sqrt{s}=240$ GeV and an integrated luminosity of $L=20~\mathrm{ab}^{-1}$. We focus on the decay modes $S\to\tau^+\tau^-$ and $S\to b\bar b$. To maximize sensitivity, we employ the particle-level transformer networks Particle Transformer (ParT) and its more-interactive variant MIParT, which exploit the features of all reconstructed objects and their correlations. For a representative signal benchmark, this approach improves the expected statistical precision on the signal strength by factors of 2.3 in the $\tau^+\tau^-$ channel and 1.4 in the $b\bar b$ channel compared to a cut-based analysis. Within the flipped Next-to-Two-Higgs-Doublet Model (N2HDM-F), the CEPC can measure the signal strength with a statistical precision down to 1.0% in the $\tau^+\tau^-$ channel and 0.69% in the $b\bar b$ channel using MIParT. It can achieve a $5\sigma$ discovery for $\mu_{\tau\tau}^{ZS}>1.6\times10^{-2}$ or $\mu_{bb}^{ZS}>5.0\times10^{-3}$, and reach 1% precision for $\mu_{\tau\tau}^{ZS}>0.93$ or $\mu_{bb}^{ZS}>0.14$. These gains are expected to qualitatively carry over to other future lepton colliders such as FCC-ee and the ILC. Our results demonstrate the potential of particle-level machine-learning techniques to strengthen light Higgs searches at future $e^+e^-$ Higgs factories.