Deep Neural Networks for Heavy Lepton-Flavor-Violating Higgs Searches at the LHC

We study lepton-flavor-violating (LFV) decays of a heavy Higgs boson, $H \to \mu\tau$, in the Type-III two-Higgs-doublet model by recasting the CMS search at $\sqrt{s} = 13$ TeV with 35.9 fb$^{-1}$ using fast detector simulation in the mass range 200-450 GeV. We develop a deep neural network (DNN) classifier trained on final-state kinematic variables that, with mass-dependent threshold optimization, reduces the expected 95% CL upper limits on the signal cross section by 42-46% in the 0-jet channel and 36-40% in the 1-jet channel relative to the standard collinear mass ($M_\mathrm{col}$) baseline. We apply SHAP interpretability analysis to identify the visible mass $m_\mathrm{vis}$ as one of the dominant discriminating feature, reflecting the characteristic neutrino momentum fraction of the $\tau$ decay. We show that supplementing the $M_\mathrm{col}$ analysis with a simplified mass-dependent pre-selection, $m_\mathrm{vis} < f \cdot m_H$ with $f = 0.7$ (0-jet) and $f = 0.8$ (1-jet), consistently improves the sensitivity over the $M_\mathrm{col}$-only baseline without requiring multivariate infrastructure. In addition, a DNN regression model trained to predict the ratio $m_H/M_\mathrm{col}$ corrects the systematic prediction bias inherent in the collinear approximation, maintaining an absolute mass prediction error below 1 GeV for signals up to 400 GeV and improving the mass resolution by 12% (0-jet) and 21% (1-jet) at $m_H = 450$ GeV. These results demonstrate a clear path toward significantly enhanced sensitivity in LFV Higgs searches at the LHC.