Title Page
Abstract
Contents
Chapter 1. Introduction 12
Chapter 2. Theoretical Foundations 15
2.1. The Standard Model of Particle Physics 15
2.1.1. The Top Quark 17
2.1.2. The CKM Matrix 18
2.2. Collider Physics 19
Chapter 3. Machine Learning 21
3.1. Decision Tree 21
3.2. Artificial Neural Networks 23
3.2.1. Feedforward Neural Network 24
3.2.2. Transformer 25
3.2.3. Deep Learning Appliance in High Energy Physics 28
Chapter 4. Event Generation and Simulation 30
4.1. Event Generation 30
4.2. Detector Response Simulation 31
4.2.1. The CMS Detector 32
4.2.2. Delphes 34
Chapter 5. Analysis Strategy 39
5.1. Object Selection 39
5.2. Event Selection 43
5.3. t → s Jet Tagging Strategy 44
5.4. Applying Machine Learning Method 48
5.4.1. The BDT Model Training 48
5.4.2. The Attention-Based Model Training 49
5.4.3. Training Variables 50
5.5. Statistical Inference for the Signal 53
5.5.1. Log-likelihood Fit 55
5.5.2. Expected Significance 55
Chapter 6. Results 61
6.1. Deep Learning Model Selection 61
6.1.1. Model Selection 61
6.1.2. Model Response 63
6.2. Statistical Inference Results 64
Chapter 7. Conclusion 67
References 69
국문초록 78
Table 2.1. The magnitude of the CKM matrix elements. Values come from the global fit of elements. 18
Table 4.1. Cross-section of tt, Drell-Yan, single top, and dibosons calculated with MADGRAPH5_aMC@NLO.[이미지참조] 32
Table 4.2. Energy deposit fraction of long-lived particles in ECAL (fECAL), and HCAL (fHCAL)[이미지참조] 36
Table 5.1. Event yields for each dilepton channel after selection criteria. 44
Table 5.2. A simple diagram of the model input and output structure for an event. J, L, and pTmiss refer to the sequence of the input variables of jet,...[이미지참조] 49
Table 5.3. Input features for the models. The BDT model only uses jet variables. The model using the jet variables uses the whole variables in this table.[이미지참조] 51
Table 5.4. Jet constituent encoder input features. 53
Table 6.1. Hyperparameter list 62
Figure 2.1. The elementary particles of the standard model. The first three columns list the three generations of fermions (quarks and lepton). The right... 16
Figure 2.2. Detector coordination system. 20
Figure 3.1. Representation of a decision tree[20]. Blue rectangles are nodes that perform decisions based on their criterion. Leaves are located at a ter-... 22
Figure 3.2. Feed-forward network flow 24
Figure 3.3. Graph of multi-head attention 26
Figure 3.4. Model architecture of the transformer 27
Figure 3.5. Self Attention for Jet Assignment model architecture[13]. Jet-wise feed-forward network (left), Multi-head self-attention block (middle),... 29
Figure 4.1. Cutaway diagram of CMS detector 33
Figure 4.2. Slice of the CMS detector[37]. It simply shows which particles are measured by which detector. 34
Figure 4.3. Tracking efficiencies for the DELPHES simulation input 35
Figure 4.4. Typical DELPHES fast simulation's work-flow diagram[43]. Event inputs coming from previous Monte-Carlo generations (MADGRAPH,... 38
Figure 5.1. Top left (right) plot shows transverse momentum of the leading(subleading) lepton. Bottom left (right) plot shows pseudo-rapidity of the... 40
Figure 5.2. Top left (right) plot shows transverse momentum of the leading(subleading) lepton. Bottom left (right) plot shows pseudo-rapidity of the... 41
Figure 5.3. Top left (right) plot shows transverse momentum of the leading(subleading) lepton. Bottom left (right) plot shows pseudo-rapidity of the... 42
Figure 5.4. Di-lepton mass distribution of ee (left), eμ (middle), μμ (right) channels. The dashed area indicates the statistical uncertainty. 43
Figure 5.5. Blocks used in SAJA-Dilepton models. Feed-forward block (left), Self-Attention block (middle), and Decoder Block (right). 45
Figure 5.6. The architecture of SAJA-Dilepton (Left). This network takes jet high-level features, lepton features, and pTmiss features. The jet constituent...[이미지참조] 47
Figure 5.7. Distribution of momentum components of jets. The signal is tt→ sWbW channel, and the background is tt → bWbW channel.[이미지참조] 56
Figure 5.8. Distribution of number of particles in jets. The signal is tt →sWbW channel, and the background is tt → bWbW channel.[이미지참조] 57
Figure 5.9. Distribution of jet shape, fragmentation, b tagging flag, and jet charge. The signal is tt → sWbW channel, and the background is tt →...[이미지참조] 58
Figure 5.10. Distribution of lepton features. The signal is tt → sWbW channel,and the background is tt → bWbW channel.[이미지참조] 59
Figure 5.11. Distribution of missing transverse momentum. The signal is tt→ sWbW channel, and the background is tt → bWbW channel.[이미지참조] 60
Figure 6.1. Learning curve of jet constituents model. 63
Figure 6.2. t → s score distributions for tt → sWbW (red), tt → bWbW(blue), Z/γ → ℓℓ (green), and Single Top (orange). The event generates...[이미지참조] 63
Figure 6.3. The separation power of the signal process from the three most contributing background processes (tt →bWbW (left), Single Top (middle),...[이미지참조] 64
Figure 6.4. Expected limits of each model with the LHC Run 2 Luminosity(137.6 fb¯¹). SAJA using the jet constituent model (Green) can best limit... 65
Figure 6.5. Expected significance with test static with zero signal strength to exclude scenarios with |Vts|=0. The significance is calculated from...[이미지참조] 65