Experimental Determination of BSM Triple Higgs Couplings at the HL-LHC with Neural Networks

Abstract: The shape of the Higgs potential is modified by the presence of additional scalar fields, as predicted in many Beyond-Standard-Model (BSM) scenarios. In such cases, deviations in the Higgs self-interactions, in particular the trilinear Higgs couplings, could serve to disentangle the physics beyond the Standard Model (SM). While the SM predicts only one trilinear Higgs coupling, extended scalar sectors allow for additional self-interactions that can manifest themselves in Higgs pair production, via the $s$-channel contribution of a heavy $\mathcal{CP}$-even scalar $H$. We present the first sensitivity study to such a BSM trilinear scalar coupling using machine learning. Specifically, we train a neural network on the invariant mass distributions of Higgs pair production at the HL-LHC to extract $\xi_H^t \times \lambda_{hhH}$, i.e. the product of the resonant $H$ top-Yukawa coupling and the trilinear coupling of $H$ to the two SM-like Higgses in the final state, $hh$. Assuming a hypothetical $H$ mass of 450 GeV, we show that, depending on future experimental efficiencies and uncertainties, a determination of $\xi_H^t \times \lambda_{hhH}$ at the 10-20% level may be achievable by the end of the HL-LHC. We present a simple and more efficient alternative to classical statistical methods, proving the efficiency of neural networks for both hypothesis testing and parameter estimation, which outperforms conventional maximum likelihood methods in this context.