Preliminary Results in the Run 3 HH → bbγγ Analysis with the CMS Experiment: Developing a ttH-Killer

Higgs boson pair production (HH or di-Higgs) in the Standard Model (SM) provides unique sensitivity to explore the structure of the Higgs potential. Measurements of its production cross section allow us to directly access the tri-linear Higgs self-coupling, λ_HHH . At the LHC, the bbγγ channel is particularly significant for non-resonant HH searches due to its clean diphoton signature, which is easy to trigger on and offers good mass resolution. At the HL-LHC, this channel alone is expected to achieve an upper limit of 1×SM [A. Tumasyan et. al. 2022].

One of the main irreducible backgrounds in HH → bbγγ is ttH → γγ due to having the same final state decay particles and an order-of-magnitude higher cross-section as compared to HH → bbγγ. To reduce the ttH → γγ background, we developed a Machine Learning (ML) model that discriminates between HH signal and ttH background events. The model (called ttH-Killer ) is heterogeneous – combining a Recursive Neural Network (RNN) with a Deep Neural Network (DNN) to better suit our data. The classifier’s preliminary performance matches its Run 2 counerpart, and we are working on improving performance using modern ML techniques. This talk will review the background and impetus for the ttH-Killer and describe the model’s performance and areas for improvement.