Variational autoencoders for regression: recovering fully leptonic $$ b\overline{b}{W}^{+}{W}^{-} $$ in di-Higgs searches

Abstract The search for double Higgs production in $$ b\overline{b}{W}^{+}{W}^{-} $$ b b ¯ W + W − , where both W bosons decay to leptons, has been rehabilitated as a good option to investigate key processes within the Standard Model scalar sector at the LHC. The missing neutrinos, however, hinder the reconstruction of useful information like the Higgs pair mass, which is very sensitive to the trilinear Higgs self-coupling. We present a solution to that problem using a Variational Autoencoder for Regression (VAER) to reconstruct the Higgs and top pairs decays hh, $$ t\overline{t}\to b\overline{b}{W}^{+}{W}^{-}\to b\overline{b}{\ell}^{+}{\ell}^{\prime -}{\nu}_{\ell }{\overline{\nu}}_{\ell^{\prime }} $$ t t ¯ → b b ¯ W + W − → b b ¯ ℓ + ℓ ′ − ν ℓ ν ¯ ℓ ′ . The algorithm predicts the invariant mass of non-resonant hh independently of specific trilinear coupling values, even for events where the Higgs self-coupling were not included during its training. VAER is also able to identify a new Higgs resonance in an unsupervised way, showing generalization power for events not presented in its training phase. Finally, we demonstrate that VAER prediction is as useful to statistical inference as ground truth simulated distributions by computing a χ2 between trilinear coupling hypotheses based on binned invariant mass distributions of $$ b\overline{b}{\ell}^{+}{\ell}^{\prime -}{\nu}_{\ell }{\overline{\nu}}_{\ell^{\prime }} $$ b b ¯ ℓ + ℓ ′ − ν ℓ ν ¯ ℓ ′ .