Speaker
Description
This study investigates the separation of the rare tH(bb) signal from background processes using machine-learning methods in the ATLAS Release 22 framework. The tH(bb) process is sensitive to the interaction between the Higgs boson and the top quark, but its small production rate makes it
difficult to distinguish from dominant background processes.
As a reference study, the previous Release 21 analysis was reproduced. The main part of the study focuses on the Release 22 analysis, where XGBoost, a graph neural network, a Graph Attention Transformer, and a multi-class Graph Attention Transformer were implemented and evaluated. The models were compared using the area under the ROC curve and the expected median limit obtained with TRExFitter.
The best result was achieved with the multi-class Graph Attention transformer, reaching an AUC of 0.91 and an expected median limit of μ = 6.05. Compared with the previous Release 22 study, which obtained an AUC of 0.82 and an expected median limit of μ = 22.4 using TabNet, this represents an improvement in expected sensitivity. We also approximated Run-3 sensitivity projection by scaling luminosity to 500 fb⁻¹, which decreased the expected median limit to 3.55.