Separation of the gluino–gluino production signal from the standard model background using deep learning

In this thesis, deep learning techniques were employed to distinguish Beyond the Standard Model (BSM) signals from Standard Model (SM) backgrounds. These methods were chosen due to the limitations of traditional analyses in handling high-dimensional and complex particle physics data. While similar tasks could be performed using statistical or rule-based methods, such approaches are generally constrained by lower accuracy and generalization capability. On simulated gluino pair production data, DNN, CNN, TabNet, and Autoencoder+TabNet models were implemented and optimized using Optuna. The results showed that the hybrid structure achieved approximately 3–5% higher performance in accuracy and AUC metrics compared to other models, with the CNN model achieving values very close to this performance. This study demonstrates that deep learning-based methods provide effective tools for classifying BSM signals and that hybrid approaches can yield significant performance improvements over individual methods.