Analysis of the ttH Process in Same-Sign Dilepton Events with Hadronic Tau Decay Using Self-Attention Architectures

The associated production of a Higgs boson with a top-antitop quark pair (ttH) is a rare but important process studied using the ATLAS detector at CERN. Separating this signal from the vast sea of background data is an impossible task for humans, and a challenging task for machines. In this work, we apply attention-based neural networks, also called transformers, to classify simulated events in the same-sign dilepton plus hadronic τ channel. We start with a simple transformer model and progress to implementing state-of-the-art techniques and concepts. Throughout the process, we analyze the inner mechanisms of the models in detail, including activation histograms and detailed headspecific attention maps. We achieve a final ROC AUC score of 0.804 and an F1-score of 0.704. Using the CERN-developed TRExFitter, we produce normalized model probability distributions to illustrate the separation between signal and background predictions and assess the stability of our classifier across different background processes. We then quantify the intrinsic statistical uncertainty in the expected t.tH coupling measurement and find that our model reduces the +/- 1σ uncertainty from μ = 1 +/- 0.28 to μ = 1 +/- 0.24.