WorksPhysics and Astronomy
Study of enhancing the signal-background separation for the Higgs boson production in association with a single top-quark in collisions at the LHC
Abstract
In this paper, a multivariate analysis based approach for enhancing the extraction of the rare Standard Model (SM) process being studied - Higgs boson production in association with a single top quark (tH) - from the prevailing background processes is presented. The analysis was performed using a Monte Carlo generated dataset simulating the outcomes of proton-proton collisions at the Large Hadron Collider (LHC). A deep fully connected neural network was developed for signal-background discrimination, incorporating a masking mechanism designed to suppress the contribution of input variables associated with objects that were not reconstructed in a given event. The performance of the masked architecture was evaluated on independent Monte Carlo samples. The results of the proposed methodology are compared with a conventional cut-based approach in terms of the statistical significance of the tH production channel. The developed framework was tested using a simulated dataset corresponding approximately to the combined statistical samples of LHC Runs 1, 2, and 3.
Introduction
The Standard Model (SM) [1, 2, 3] is the most comprehensive theory describing the fundamental interactions of elementary particles: quarks, leptons, and bosons. The construction of this theory was completed by discovering the Higgs boson [4, 5] both in the ATLAS and CMS experiments [6, 7] in 2012. However, deviations between theoretical predictions and experimental data continue to promote more precise measurements of its parameters. Moreover, unresolved issues drive scientists to search for new physical phenomena beyond the Standard Model (BSM).
Precise measurement of the Higgs boson parameters is of particular interest in elementary particle physics. Despite its experimental confirmation, some of its properties, such as the possible presence of CP symmetry violation [8–10] or the detailed structure of the Higgs potential, remain not fully explored. Specifically, the Higgs potential is sensitive to radiative corrections [11, 12], in which the top Yukawa coupling provides a dominant contribution, thereby affecting the stability of the electroweak (EW) vacuum [13]. In this context, a more detailed investigation of the Higgs boson’s interactions holds significant potential for uncovering signs of “new physics” [12] beyond the Standard Model.
To date, Higgs boson production has been observed in four channels: two production channels alongside other elementary particles (Higgs strahlung and Top fusion), and two where a single Higgs boson is produced alone (Gluon-Gluon Fusion and Vector Boson Fusion). Among all other hypothetical Higgs boson production channels, the most intriguing is its production in association with a single top quark () [14–16]. The first two leading-order perturbation theory Feynman diagrams for () process are shown in Figure 1.
Figure 1. Feynman diagrams of leading contributions for signal Higgs boson production in association with a single top-quark, followed by the Higgs boson decay into a pair and top-quark decay via leptonic mode.
The magnitude of the top-Higgs Yukawa coupling has already been indirectly constrained with a precision of approximately at the LHC [17, 18]. Direct measurements of are provided by the production process, whereas the process is directly sensitive to the relative complex phase between the and couplings through the interference of the corresponding production amplitudes [19–23]. In the Standard Model, this interference is destructive, suppressing the production rate, while a change in the relative phase can lead to constructive interference. Consequently, the process provides a unique probe of the relative phase of the top-Higgs Yukawa coupling, while statistically significant evidence for its production has yet to be established.
Being of great interest for experimental confirmation and study [14], this production channel is particularly challenging to detect because the Standard Model predicts the smallest production cross-section for (see Figure 2). In this paper, the potential for observing the production process is studied in proton-proton collisions at a centre-of-mass energy of TeV, assuming an integrated luminosity of fb. Destructive interference leads to the suppression of the total interaction cross-section to 69.66 fb at TeV [21], resulting in an extremely small signal yield compared with the background.
Figure 2. Dependence of cross-section for various Standard Model Higgs boson productions on the center-of-mass energy of proton-proton collisions [24].
Background processes comprise a multitude of possible events occurring in proton–proton collisions, distinct from the Higgs boson production in association with a single top quark (), which is the primary focus of this study. A large fraction of these background contributions can be suppressed by exploiting the presence of specific physics objects in the final-state signature. The most reasonable approach to extract the events is to select the decay modes which ensure simultaneously the highest signal statistics and suppression of the majority of background channels. For this reason, when defining the signal process signature, special attention was given to the top quark decay via the semileptonic mode (), which accounts for 33.2% of all top quark decays, and the Higgs boson decay through its dominant mode (), which occurs in 58% of cases. The signal process is characterized by the presence of at least three -jets, a charged lepton with high transverse momentum and significant missing transverse energy carried away by a neutrino. An additional “forward jet”, originating from the spectator quark , is also characteristic of the production topology (see Figure 1) [25].
However, there exists a class of background processes whose final-state signatures exactly mimic that of the signal. Consequently, further discrimination based solely on object-level selection becomes insufficient, necessitating the use of multidimensional kinematic information to enhance the signal-to-background separation. This issue is further exacerbated by the fact that at 13.6 TeV the () production cross-section is four orders of magnitude smaller than the main background process () with 811 pb of cross-section [26, 27]. As a result, signal events are vastly outnumbered by the dominant background processes, making their extraction from the prevailing and kinematically similar background particularly challenging.
The outcome of each collision event is characterized by a finite set of kinematic observables that describe the corresponding reconstructed physical objects in the final state. Each event can therefore be represented by a feature vector , where denotes the number of variables used. The collection of such features forms a -dimensional phase space in which each event corresponds to a single point . Signal and background events are distributed in this space according to the probability density functions and , respectively.
Intuitively, the so-called signal region can be identified in those regions of the feature space where the probability of observing a signal event exceeds that of observing a background event, i.e. . In general, the signal and background regions do not necessarily form single connected domains with clear boundaries. Instead, since there may exist regions in which the probability densities of the corresponding distributions coincide, the signal-enriched regions may overlap with the background-enriched regions. Furthermore, the complex nonlinear structure of the signal probability density in phase space significantly limits the effectiveness of conventional cut-based approaches for optimal signal region definition beyond a certain stage of the event selection procedure.
Therefore, the signal–background classification problem is inherently statistical in nature and consists in constructing a decision rule that minimizes the probability of classification error. From the perspective of statistical decision theory, the optimal classification rule is determined by the Neyman–Pearson lemma. According to this lemma, the most powerful test for discriminating between the signal and background hypotheses is based on the likelihood ratio [28]:
Event classification is performed by introducing a threshold value , which defines a separating hypersurface in the feature space corresponding to the boundary . In this case, the decision rule can be written as:
The choice of the threshold value is determined according to a selected statistical criterion in order to maximize the sensitivity to the signal while minimizing the statistical uncertainty.
In practice, obtaining explicit analytical expressions for the probability density functions and is generally infeasible, since their evaluation involves multidimensional phase-space integration of matrix elements including higher-order QCD contributions. Moreover, due to the similarity of the final state signatures of signal and some background events in the collider, the true origin of each recorded event (signal or background) is unknown. This precludes the direct use of experimental data for reconstructing these distributions. Instead, in order to obtain labeled datasets suitable for further analysis, specialized event generators are employed, which incorporate up-to-date theoretical models of elementary particle physics. These labeled datasets provide an approximate representation of the real distributions in the feature space.
In this study, we present the development of a methodology aimed at improving the efficiency of identifying the rare process (). The analysis is performed using Monte-Carlo generated datasets that include both signal and background processes.
Achieving the highest possible signal region determination accuracy will facilitate reaching the signal significance level sufficient for an observation and confirmation of the production channel at the earliest possible stage. Furthermore, in future measurements of various physical properties in this channel, achieving the best possible selection accuracy will also improve the precision of estimating the corresponding physical parameters.
Monte-Carlo samples
The signal and most background processes ( [29], [30], [31], [32], [33]) were simulated using the MadGraph5_aMC@NLO [34] Monte Carlo (MC) generator, while parton showering and hadronization were modeled with the Pythia8 [35]. The leading-order Feynman diagrams for all considered background processes are presented in Figure 7. The main background process () was simulated using POWHEG [36] and Herwig7 [37] as the parton shower generator. The matrix elements for signal and background processes with the largest cross-section (, ) were calculated in the next-to-leading order of perturbation theory. Jet clustering was performed using the FastJet package [38]. The following kinematic selection criteria have been applied to isolate the signal reactions (see Figure 1 and § 1) [25]:
Figure 7. Leading channels for each background process. Where each Feynman diagram shows the first order of perturbation theory for , , , , , and processes, respectively.
Transverse momentum () for the leading lepton (the charged lepton with the highest ) of at least 25 GeV/c. This restriction corresponds to one of the triggers for recording a charged lepton event in the ATLAS experiment.
, where is the pseudorapidity of the leading lepton; accounting for the geometry of the ATLAS and CMS detectors.
GeV, where is the transverse momentum of the reconstructed neutrino; the missing transverse energy is reconstructed in the ATLAS and CMS experiments with a resolution of around 10 GeV.
, where is the pseudorapidity of the b-jets; a limit roughly corresponding to the acceptance of the ATLAS [39, 40] and CMS [41–43] tracking systems.
Table 1 provides information on the number of events simulated and selected for inclusion in the database for signal and background processes.
| Process | ME order | (fb) | Preselection yield | ||
| NLO | 13.4 | 822 | |||
| NLO | |||||
| LO | |||||
| NLO | 129 | ||||
| LO | 34.1 | ||||
| LO | 31.4 | ||||
| LO | 212 | 792 | |||
| Total background |
Table 1. Table 1. Statistics of simulated and selected events in this study. is the number of generated events; is the process cross-section; is the number of generated events passing the selection described above; The preselection yield is the expected number of events after the preliminary cuts, normalized to an integrated luminosity of 500 using the process cross-sections and the selection efficiencies estimated from simulation.
The neutrino transverse momentum () was reconstructed as the vector opposite to the total transverse momentum of all jets and charged leptons. To determine the longitudinal component of the neutrino momentum, the invariant mass of the leading lepton and the neutrino was set equal to the nominal mass of the W-boson ( GeV). This constraint leads to two solutions:
where .
The jet clustering was performed using the anti- algorithm [44] with the following parameters: minimum jet transverse momentum GeV, jet radius parameter , and pseudorapidity range for all jets .
A jet was tagged as a “b-jet” only if the distance between the b-quark momentum and the jet momentum satisfied and the ratio of the jet energy to the b-quark energy was in the range .
The main background process for the studied Higgs boson production channel is ; given the requirement of one charged lepton, the decay channel takes the form . To first approximation, the background process can be suppressed by requiring at least three b-jets in the event. Due to the limited precision of experimental detectors, a c-quark jet can be misidentified as a b-jet with a probability of about 12% in the ATLAS [45] and CMS [46] experiments. Jets from light quarks are misidentified as b-jets with a probability of about 0.3% [45, 46]. This effect was taken into account by assigning a b-tag to random jets from c and quarks with corresponding probability when simulating the process.
Event reconstruction requires an assignment of every b-jet to the decay of either top quark or Higgs boson. This is performed by selecting the combination that provides the best agreement between the reconstructed masses of the top quark and the Higgs boson and their nominal values. The top quark mass is reconstructed from the leading lepton, neutrino, and a b-jet, while the Higgs boson mass is reconstructed from the remaining two b-jets. In the case when four b-jets were reconstructed in the event, three of them with the highest were selected for the reconstruction of the top-quark and Higgs boson masses.
From the three b-jets and the two solutions for the neutrino momentum, six combinations can be formed to reconstruct the top quark and Higgs boson masses. For each combination, a value was calculated using the formula:
where the parameters , , , were determined by fitting the distributions for the reconstructed masses corresponding to the correct combination. The reconstructed masses were determined from the combination corresponding to the smallest . The distributions of the invariant mass of the Higgs boson and the top-quark, shown in Figure 3, demonstrate the difficulty of separating the signal from the total background using rectangular cuts. The Separation Power [47], defined in equation (5), was used to quantify the difference between the signal and background distributions of an individual variable over its full range. The signal and background histograms were constructed using event weights that account for the process cross-sections, the integrated luminosity, and the sign of the generator weight for each event.

Figure 3. The graph on the right illustrates the invariant mass of the reconstructed Higgs boson, while the one on the left illustrates a top-quark for the combination with the smallest .
where is the number of histogram bins, and and denote the weighted number of events for signal and background in the -th bin, respectively. Larger values of indicate greater discriminating power.
In the process, a “spectator quark” is produced (Figure 1), whose momentum is predominantly directed towards regions of high absolute pseudorapidity (typically ). The jet originating from this quark can be used to discriminate the signal from the main background process , where the forward jet production is not preferred [25]. We tested two approaches to identify the jet from the spectator quark: selection of the “forward jet” with GeV/c, , and selection of the “tagging jet”, i.e. a non- jet having the largest invariant mass with the -jet from the top-quark decay [48]. The pseudorapidity distributions of the “forward” jet selected by the two algorithms are presented in Figure 4. The use of the tagging jet provides a better discrimination between signal and background events, as is evident from the distributions shown in Figure 5-Figure 6.
Figure 4. Pseudorapidity of the forward jet (left panel) and the tagging jet (right panel).
Figure 5. Distance in between top-quark and forward jet (left panel), top-quark and tagging jet (right panel).
Figure 6. Invariant mass of top-quark and forward jet (left panel), top-quark and tagging jet (right panel).
Fifty kinematic variables were constructed and considered for the further training of the neural network that separates signal from background. The full list and description of the kinematic variables are provided in appendix A.
Exploratory data analysis
Prior to supplying the input features into the machine learning model, several preprocessing steps were performed. The preprocessing stages described below were applied in a specified sequential order to preserve consistency of the data transformation procedure.
Correlation analysis
First, a feature Correlation Matrix (see Figure 8) was constructed, which allowed the identification of five pairs of variables with the Pearson correlation coefficient [49, 50] greater than or equal to . Pairs of highly linearly correlated variables are shown in Table 2. From each pair of strongly correlated variables only one feature was retained - specifically the one exhibiting the higher Separation Power (see equation (5)).

Figure 8. Correlation matrix for total events.
| № | Variable 1 | Variable 2 | Pearson coefficient |
| 1 | n_nonbjet | n_jet* | 0.971 |
| 2 | sphericity_all_jets | sphericity_lnu* | 0.879 |
| 3 | pt_tag | max_pt_nonbjet* | 0.866 |
| 4 | aplanarity_all_jets* | aplanarity_lnu | 0.812 |
| 5 | pt_higgs* | max_pt_bjet | 0.807 |
Table 2. Pairs of highly correlated variables. Variables marked with an asterisk (*) are excluded from the final input feature set.
Normalization
To promote proportionate contributions of all input variables during training, feature normalization was implemented. Normalization refers to the linear transformation of a variable such that its values are mapped onto a fixed interval , while preserving the overall shape of its distribution.
Dataset splitting
The initial dataset was partitioned into training (70%), validation (15%), and testing (15%) subsets. A stratified split was employed based on the process label of each event to preserve the relative proportions of all individual processes across all subsets.
Development and optimization of the neural network
Mathematical formulation of the problem
Given a finite set of processed variables, the objective of this study could be formulated as the reconstruction of an accurate approximation of the likelihood ratio based on simulated labeled events. Although the reconstructed likelihood ratio could, in principle, be obtained through independent estimation of the signal and background probability density functions, and , respectively, such density estimation becomes statistically intractable in high-dimensional feature spaces due to the curse of dimensionality. The number of events required for a reliable estimation of a probability density in a -dimensional space scales exponentially as , where denotes the effective binning granularity per dimension. Consequently, direct density reconstruction becomes impractical due to the unrealistically large event statistics required.
Instead of estimating the likelihood ratio directly, it is often more accurate and convenient to express the classification problem in terms of the posterior probability that an event with observed kinematic features originates from the signal process. These two quantities are related through Bayes’ theorem as
where and denote the prior probabilities of the signal and background processes, respectively. From this relation it follows that the posterior probability and the likelihood ratio represent monotonic transformations of one another. Therefore, both quantities contain equivalent information from the perspective of the classification problem. However, the direct approximation of the posterior probability constitutes a statistically more stable learning task.
To address this problem, a machine learning model with strong nonlinear approximation capabilities was employed, namely a deep neural network [51–53]. From a mathematical perspective, a neural network represents a parameterized nonlinear mapping , where denotes the set of trainable model parameters (weights and biases). Training consists of iteratively adjusting the parameters in order to minimize a loss function defined on the training dataset. For supervised training with the backpropagation algorithm, each event is assigned a binary target label: for signal events, while is assigned to background events.
With a sigmoid activation function in the output neuron , the natural choice of loss function is the binary cross-entropy. In the idealized limit, the function minimizing the expected binary cross-entropy risk coincides with the true posterior probability of the signal class, i.e. [54].
During training, event weights accounting for the process cross-sections and integrated luminosity were applied, followed by equalization of the signal and total background yields. Generator-level event weights were not incorporated, in order to avoid negative weights in the training sample and to prevent the associated risk of the loss function from becoming unbounded from below [55]. Under this balanced training, the network response also becomes a monotonic function of the likelihood ratio between the effective signal and background distributions. The trained network therefore defines a scalar field over the phase space, which was used as a nonlinear discriminant to extract the signal events.
Masking
In some events, individual physical objects or categories may be absent. In such cases, the values of the corresponding kinematic variables are undefined. To preserve a fixed input dimensionality of the neural network, all missing values were encoded with a common placeholder value chosen well outside the physically allowed range of the corresponding variables.
To prevent the network from treating the artificial placeholder value as a numerical feature and thereby learning unphysical correlations induced by it, a dedicated mechanism for masking missing values was implemented. The implementation of this approach requires the network input to be first separated into isolated branches. Each branch is associated with a specific object or category and consists of a small fully connected subnetwork. Its input is a set of kinematic variables describing the corresponding entity. This organization allows different entities to be processed independently and the masking to be applied at the level of the entire input group under consideration. This is important because missing values of variables belonging to a given branch occur simultaneously in the same events.
To conditionally control the contribution of a branch depending on the presence of the corresponding entity, the transformation of its hidden representation was modified according to equation (7).
here denotes the input vector of the branch, and are the weight matrix and bias vector of its fully connected layer, respectively, and denotes the activation function. The quantities and are binary masks whose dimensions correspond to the input vector and the hidden representation of the branch, respectively. If at least one value in the input group is equal to the placeholder value, all elements of both masks are set to zero. If the input group contains physically defined values, all elements of both masks are set to one.
This construction provides two important properties of the specialized branch when missing values are supplied to it:
Formation of a zero-valued hidden representation of the branch.
Absence of updates to the branch weights and biases due to the corresponding gradients from the loss function being zeroed.
Taken together, these properties prevent kinematic information associated with an absent object from being propagated to the subsequent classification transformation, while such an event does not contribute gradients for updating the parameters of the corresponding branch. When physically defined values are provided, both and consist entirely of ones, and equation (7) reduces to the standard transformation of a fully connected branch. Thus, for present objects, the masking mechanism does not alter the standard computations of the network.
Each branch produces its own hidden representation of the corresponding input group. The resulting representations are concatenated into a single vector. This vector is further augmented with a number of binary presence features equal to the number of masked branches, provided separately from the hidden kinematic representations. Each of these features characterizes the presence of the corresponding entity in the event. A binary presence feature is dynamically constructed for each masked branch: a value of 1 is assigned if the entire input group has physically defined values, and a value of 0 if all input variables take the placeholder value. The explicit provision of this information supplies the classifier with additional discriminating information independent of the corresponding kinematic representation.
The resulting vector is provided as the input to the subsequent common classifier, which is implemented as a fully connected deep neural network. The overall architecture of the proposed neural network is illustrated in Figure 9.

Figure 9. General neural network architecture used in the analysis, incorporating the masking mechanism.
Optimization
According to the universal approximation theorem [56, 57], a feedforward neural network with at least one hidden layer (perceptron) and a sufficient number of neurons is capable of approximating an arbitrary continuous function on a compact set with a predetermined accuracy. From this perspective, one may assume that almost any network architecture, provided that the hidden layers contain a sufficiently large number of neurons, can achieve the required expressive power.
However, in practice the effectiveness of such approximation is limited by imperfections of the training process, such as the possibility of getting convergence to local minima of the loss function, the risk of overfitting, statistical noise in the data, and limitations of the regularization algorithm. Consequently, training an arbitrary architecture does not guarantee convergence to the global minimum of the loss function and therefore does not ensure an ideal approximation of the objective function as predicted by theory.
Due to the analytically unpredictable influence of the architecture on the geometry of the loss landscape, it is not possible to determine in advance the optimal combination of architectural and training parameters that provides the best approximation quality for the target dependence. To identify the most suitable configuration, an automated hyperparameter optimization procedure [58, 59] was performed using the Optuna library [60]. The optimization involves a systematic exploration of all possible predefined combinations of hyperparameters within specified ranges.
An additional step aimed at reducing the computational cost of the optimization procedure was the use of a subsample of events selected from the full dataset that passed the preliminary event selection. The reliability of this approach has been empirically demonstrated in [61]. After the initial partitioning (see § 3), a reduced dataset was independently selected from each subset by retaining a fixed number of events for each process. The composition of the resulting training, validation, and testing datasets, containing the same number of events for each process, is summarized in Table 3. Additionally, the 33 variables containing the dominant discriminating information were selected using the permutation importance method [62] for the hyperparameter optimization procedure.
| Process | Number of events |
| 50,000 | |
| 20,000 | |
| 10,000 | |
| 6,000 | |
| 6,000 | |
| 6,000 | |
| 3,000 |
Table 3. Substatistics of processes used for the hyperparameter optimization process. The size of the subsample for each process was chosen such that the actual number of signal events was approximately equal to the total number of background events.
Since the maximum achievable signal significance depends on the selection threshold, its evaluation relies on only a fraction of the available event statistics and is therefore sensitive to statistical fluctuations. To objectively evaluate the performance of each candidate architecture and guide the optimization toward the highest achievable accuracy, the area under the receiver operating characteristic curve (AUC) [63] was chosen as the target metric. By exploiting the full available signal and background samples, AUC provides a more stable basis for comparing a large number of network architectures.
A Bayesian exploration [64–66] of all possible combinations was employed in this study — the Tree-structured Parzen Estimator (TPE) algorithm [67]. The convergence dynamics of the optimization process is shown in Figure 10.
Figure 10. Stepwise convergence of the neural network hyperparameter optimization, showing the best score achieved as a function of the architecture index.
Among all 1300 configurations tested, the best value of the target metric was achieved at the 1009th architecture. Following this point, the algorithm continued exploring the local region of the parameter grid for an additional 291 iterations. However, no further improvement in the classification performance was observed, indicating that the optimization process had reached a saturation stage. Any potential further improvements in classification fication accuracy would likely be negligible compared to the accuracy already achieved, given the additional computational resources required. Therefore, it was concluded that terminating the optimization procedure at this stage was justified.
Final model evaluation
The architecture and training parameters of the best-performing neural network identified by the hyperparameter optimization are summarized in Table 4. The final training of this architecture was performed using the full available event statistics and the complete set of 41 input kinematic variables. The input variables are described in Table 5 in the same order as they are provided to the neural network. The same event weighting scheme as described in § 2 for the calculation of the separation power metric 2.3 was used to construct the evaluation plots.
| Branches | ||||
| Name | Input neurons | Activation | Dropout | Hidden neurons |
| always_filled | 24 | tanh | 0.4 | 128 |
| w_qq | 2 | mish | 0.2 | 6 |
| forward_jet | 7 | mish | 0.7 | 12 |
| tagging_jet | 8 | swish | 0.2 | 16 |
| Common classifier | ||||
| Layer | Neurons | BatchNorm. | Activation | Dropout |
| Input layer | 165 | No | No | No |
| Fully Connected 1 | 224 | Yes | swish | 0.45 |
| Fully Connected 2 | 96 | Yes | mish | 0.2 |
| Fully Connected 3 | 192 | Yes | relu | 0.25 |
| Output Layer | 1 | No | Sigmoid | No |
| Training settings | ||||
| № | Parameter | Value | ||
| 1 | Loss function | Binary Crossentropy | ||
| 2 | Batch size | 512 | ||
| 3 | Number of epochs | 100 | ||
| 4 | Optimizer | Adam | ||
| 5 | Learning rate |
Table 4. Optimal neural network architecture and training parameters identified through optimization.
| Category | № | Encoding | Variable description | SP (%) |
| always-filled | 1 | mass_3jet | Invariant mass of three jets with the largest | 13.22 |
| always-filled | 2 | sphericity_all_jets | Measure of uniformity of the distribution of jets in the space | 14.85 |
| always-filled | 3 | aplanarity_lnu | Measure of deviation of objects from the common plane | 14.87 |
| always-filled | 4 | delta_eta_top_higgs | Difference of pseudorapidities from the top quark and Higgs boson | 6.08 |
| always-filled | 5 | mass_th | Invariant mass of top quark and Higgs boson | 4.90 |
| always-filled | 6 | mass_higgs | Reconstructed mass of Higgs boson | 3.16 |
| always-filled | 7 | mass_top | Reconstructed mass of the top quark | 2.82 |
| always-filled | 8 | q_lead | Charge of the lepton with the largest | 1.55 |
| always-filled | 9 | n_nonbjet | Number of jets produced by quarks other than a quark | 1.34 |
| always-filled | 10 | fwjets2 | Second Fox–Wolfram moment composed of the momenta of jets | 4.64 |
| always-filled | 11 | fwlnujet1 | First Fox–Wolfram moment composed of the momenta of objects | 1.56 |
| always-filled | 12 | mT_all_jets | Transverse mass of all jets | 11.83 |
| always-filled | 13 | eta_lead | Pseudorapidity of the lepton with the largest | 0.03 |
| always-filled | 14 | sum_pt_jets | Algebraic sum of transverse momenta of all jets | 1.27 |
| always-filled | 15 | delta_phi_higgs_top | Difference between the azimuthal angles of the top quark and Higgs boson | 0.92 |
| always-filled | 16 | max_pt_bjet | Transverse momentum of the jet with the largest | 1.09 |
| always-filled | 17 | pt_second_bjet | Transverse momentum of the jet second in its value | 0.52 |
| always-filled | 18 | eta_higgs | Reconstructed pseudorapidity of Higgs boson | 0.76 |
| always-filled | 19 | eta_second_bjet | Pseudorapidity of the jet second in its value | 0.80 |
| always-filled | 20 | m_max_bjet | Invariant mass of the jet with the largest | 0.17 |
| always-filled | 21 | eta_max_bjet | Pseudorapidity of the jet with the largest | 1.01 |
| always-filled | 22 | n_central_jet | Number of central light jets | 11.68 |
| always-filled | 23 | chi2_min | Quality ( criterion) of determining the masses of Higgs boson and top quark | 2.05 |
| always-filled | 24 | pt_lead | Largest transverse momentum among leptons | 0.39 |
| w-qq | 25 | deltaR_qq_W_max_pt | Difference between the rapidities of the light jets with the largest | 8.87 |
| w-qq | 26 | deltaR_qq_W_nearest_mass_W | Difference between the rapidities of the light jets whose masses are closest to the mass of the W boson | 4.68 |
| forward-jet | 27 | delta_eta_higgs_fwd | Difference between the pseudorapidities of Higgs boson and forward jet | 0.96 |
| forward-jet | 28 | delta_eta_top_fwd | Difference of pseudorapidities of the top quark forward jet | 4.99 |
| forward-jet | 29 | pt_fwd | Transverse momentum of the forward jet | 3.87 |
| forward-jet | 30 | eta_fwd | Forward jet pseudorapidity | 3.25 |
| forward-jet | 31 | mass_top_fwd | Invariant mass of the top quark and forward jet | 9.08 |
| forward-jet | 32 | deltaR_bjetmax_fwd | Difference between the pseudorapidities of the forward jet and the jet with the largest | 2.82 |
| forward-jet | 33 | deltaR_nearest_bjet_fwd | Difference between the rapidities of the forward jet and the jet closest to it | 2.49 |
| tagging-jet | 34 | delta_eta_higgs_tag | Difference between the pseudorapidities of Higgs boson and tagging jet | 4.05 |
| tagging-jet | 35 | pt_tag | Transverse momentum of the tagging jet | 1.87 |
| tagging-jet | 36 | eta_tag | Tagging jet pseudorapidity | 15.01 |
| tagging-jet | 37 | delta_eta_top_tag | Difference of pseudorapidities of the top quark tagging jet | 11.49 |
| tagging-jet | 38 | mass_higgs_tag | Invariant mass of Higgs boson and tagging jet | 14.71 |
| tagging-jet | 39 | mass_top_tag | Invariant mass of the top quark and tagging jet | 18.38 |
| tagging-jet | 40 | deltaR_bjetmax_tag | Difference between the pseudorapidities of the tagging jet and the jet with the largest | 9.63 |
| tagging-jet | 41 | deltaR_nearest_bjet_tag | Difference between the rapidities of the tagging jet and the jet closest to it | 10.96 |
| 42 | n-jet* | Number of jets | 0.48 | |
| 43 | sphericity_lnu* | Measure of uniformity of the distribution of objects in the space | 14.44 | |
| 44 | max_pt_nonbjet* | Largest transverse momentum among light jets | 1.39 | |
| 45 | aplanarity_all-jets* | Measure of deviation of the jets from the common plane | 14.86 | |
| 46 | pt_higgs* | Reconstructed transverse momentum of Higgs boson | 0.98 | |
| 47 | mass_higgs_central_jet† | Invariant mass of Higgs boson and central light jet | 6.45 | |
| 48 | n_forward_jet† | Number of forward light jets | 8.59 | |
| 49 | mass_higgs_fwd† | Invariant mass of Higgs boson and forward jet | 5.18 | |
| 50 | energy_third_bjet† | Energy of the jet third in its transverse momentum | 1.40 |
Table 5. Table 5. Variables description. The variable marked with an asterisk (*) are excluded due to strong correlations, while variables marked with a dagger (†) are removed based on their negative contribution, as determined by the permutation importance method.
The absence of significant overfitting was assessed based on the lack of substantial divergence between the training and validation curves during training, as shown in Figure 11. The final performance of the neural network was evaluated by visualizing its predictions through a histogram, as shown in Figure 12. The agreement between the network output distributions for the training and testing datasets, reflected in the ratios remaining close to unity over the full range of the output discriminant, further supports the absence of significant overtraining.
Figure 11. The evolution of neural network performance metrics on training and validation events as a function of the number of training epochs.
Figure 12. Distribution of neural network output predictions. The red dashed histogram represents the likelihood of the network identifying events in which a Higgs boson is produced in association with a single top quark, classifying them as signal events. In contrast, the blue filled histogram corresponds to the probability of background events. The lower panels show the bin-by-bin ratios between the training and testing samples separately for the total event set, the signal process, and the background contribution.
To evaluate the overall classification performance of the neural network, the Receiver Operating Characteristic (ROC) curve (see Figure 13) was employed. The Area Under the Curve (AUC) metric quantifies the probability that the classifier assigns a higher score to signal than to background events. The achieved classification performance reaches
Figure 13. ROC-Curve characterizes the overall discriminating power of a binary classifier.
, where the uncertainty due to the limited simulation statistics was estimated using bootstrap method [68].
Physical sensitivity assessment
The statistical signal significance [69] was determined using the Asimov approximation:
where represents the expected number of signal events and denotes the expected number of background events. This metric characterizes the statistical reliability of the observed signal with respect to background fluctuations.
The maximization of the signal significance is accomplished by introducing a threshold in the scalar output space, which defines the decision boundary . This boundary isolates the signal region , where the signal contribution is expected to be statistically most pronounced relative to the background, in accordance with equation (8).
To determine the optimal value of the neural network output threshold, the dependence of the signal significance on is studied, as shown in Figure 14. For each threshold value, the expected numbers of signal and background events falling within the signal region are evaluated. These values are then used to compute the corresponding signal significance. The associated statistical uncertainty of the signal significance was analytically evaluated using error propagation for indirect measurements.
Figure 14. Dependence of the expected statistical signal significance on the output threshold of the neural network.
To avoid signal regions dominated by limited signal statistics, the optimal selection threshold () was required to satisfy two simultaneous requirements, as specified in equation (9). If the threshold at which the signal significance reaches its maximum () did not satisfy all these conditions, the threshold was gradually relaxed until both conditions were fulfilled.
Assuming 500 fb integrated luminosity, the maximum expected signal significance value of satisfying both requirements is achieved at an optimal neural network output threshold value of . This is to be compared with the significance of observed at zero threshold, i.e. for all events passing the preselection. The application of this threshold value defines a signal region with a purity of . The resulting event yields are 113 2 for the signal and 882 114 for the background. This selection corresponds to a signal efficiency of 13.75 %, while the background rejection reaches 99.87 %.
The cut-and-count method selects a single bin in the high-score tail of the classifier predictions distribution. This bin corresponds to a narrow signal-enriched region, resulting in a loss of a fraction of the signal events outside this region. To account for the full available statistics, the complete distribution was divided into 39 uniformly spaced bins. The expected signal significance was evaluated independently in each bin using the same expression as in equation (8). The contributions from the individual bins were combined into the overall estimate of the signal significance according to the following prescription:
The application of the shape-fit method resulted in a further increase in the statistical signal significance to .
Conclusion
In this work, the expected sensitivity to the production of a Higgs boson in association with a single top quark was investigated, focusing on the decay channel and the semileptonic decay of the top quark . The analysis was performed using Monte Carlo simulated proton-proton collisions at a centre-of-mass energy of TeV, assuming an integrated luminosity of 500 fb.
The initial selection cuts on individual kinematic variables provided effective background suppression. Further signal selection was achieved using a neural network. For the first time, an object-aware masking mechanism was adapted into a deep fully connected neural network architecture for search. The input variables were preprocessed to mitigate the influence of anomalous values and improve the suitability of the input variables for network training. The architecture that most accurately approximates the data was determined through hyperparameter optimization.
The best-performing architecture achieved a classification accuracy of . Using the resulting classifier, the expected statistical only significance was estimated to be with the cut-and-count method and with the shape-based fit. The shape-fit method yields a factor of 4.7 improvement in the signal significance compared to the rectangular-cut-based preselection stage, which achieved a significance of .
Further improvements to the analysis will involve the implementation of realistic -tagging efficiencies and misidentification rates, accounting for detector response through simulation, and the incorporation of the corresponding systematic uncertainties. It is anticipated that accounting for these additional effects will lead to a significant reduction in the estimated signal significance compared with the current simplified simulation setup. This will provide a more realistic estimate of the sensitivity achievable in an actual LHC analysis.
Variables description
The full list of kinematic variables is given in Table 5.
Acknowledgments
The authors express sincere gratitude to Aleksey Bondyakov for his valuable technical support in configuring the interaction between the local workstation and the remote storage infrastructure of the Laboratory of Information Technologies. This contribution was essential to improve the data management workflow.
Valuable assistance was also provided by Viktoriia Kiseeva through discussions of modern approaches to masking missing values and by providing a demonstrative implementation of this approach, which served as the basis for its subsequent adaptation and application in the present analysis.
References
- [1]S. Weinberg, A model of leptons, Physical Review Letters 19 (1967) 1264–1266.DOI
- [2]S.L. Glashow, Partial-symmetries of weak interactions, Nuclear Physics 22 (1961) 579–588.DOI
- [3]A. Salam, Weak and electromagnetic interactions, in Selected Papers of Abdus Salam, p. 244–254, WORLD SCIENTIFIC (1994), DOI.DOI
- [4]F. Englert and R. Brout, Broken Symmetry and the Mass of Gauge Vector Mesons, Physical Review Letters 13 (1964) 321.DOI
- [5]P.W. Higgs, Broken Symmetries and the Masses of Gauge Bosons, Physical Review Letters 13 (1964) 508.DOI
- [6]ATLAS Collaboration and others, Observation of a new particle in the search for the Standard Model Higgs boson with the ATLAS detector at the LHC, Physics Letters B 716 (2012) 1.DOI
- [7]CMS Collaboration et al., Observation of a new boson at a mass of 125 GeV with the CMS experiment at the LHC, Physics Letters B 716 (2012) 30.DOI










