BSM reach of four-top production at the LHC

Many scenarios of beyond the Standard Model (BSM) physics give rise to new top-philic interactions that can be probed at proton machines such as the Large Hadron Collider through a variety of production and decay modes. On the one hand, this will enable a detailed determination of the BSM model ’ s parameters when a discovery is made and additional sensitivity in nondominant production modes can be achieved. On the other hand, the naive narrow width approximation in dominant production modes such as gluon fusion might be inadequate for some BSM parameter regions due to interference effects, effectively making less dominant production modes more relevant in such instances. In this work, we consider both these questions in the context of four top quark final states at the LHC. First, we show that the SM potential can be enhanced through the application of targeted graph neural network techniques that exploit data correlations beyond cut-and-count approaches. Second, we show that destructive interference effects that can degrade BSM sensitivity of top-philic states from gluon fusion are largely avoided by turning to four top final states. This achieves considerable exclusion potential, e.g., the two Higgs doublet model. This further motivates four top final states as sensitive tools for BSM discovery in the near future of the LHC. DOI: 10.1103/PhysRevD.108.035001


I. INTRODUCTION
The absence of evidence for new physics beyond the Standard Model (BSM) in searches at the high-energy frontier of the Large Hadron Collider (LHC) is on the one hand puzzling, as the fundamental origin of the electroweak scale and its stability remains a mystery.On the other hand, the BSM benchmarking programme leading up to the LHC data was, perhaps, overly optimistic.The absence of a BSM signal has highlighted less abundant collider processes as important tools for BSM discrimination, drawing from model-independent and theoretically motivated techniques as well as improved approaches to better exploiting model-(un)specific correlations in searches for new interactions.The resurgence of machine learning techniques applied to the LHC realm is an example of the latter; effective field theory (EFT) in its linear or non-linear formulation is an example of the former.
Particularly interesting processes along these lines are four top quark final states, with the ATLAS and CMS collaborations recently reporting evidence of SM-like production [1][2][3][4][5][6].Four top final states are relevant processes when considering EFT deformations of expected SM correlations [7][8][9], lending considerable sensitivity to large-scale EFT fits, but they also enable a targeted snapshot of electroweak Higgs properties [10].Recently in [11], it was demonstrated that the large kinematical information that can be exploited in such a busy final state leads to improved sensitivity when final state correlations are tensioned against the SM backgrounds using machine learning approaches.Reference [12] discussed Bayesian techniques for four tops.
For top-philic resonance extensions, four top final states are naively less motivated as gluon fusion provides a large production cross section and top pair resonance searches are experimentally less challenging than the analysis of four top states.Given the generic vulnerability of pp → t t final states to quantum interference with QCD backgrounds [13,14], this might well be a fallacy: when BSM signals are rendered small through interference in the dominant production and decay channels, only those with limited production cross sections remain.Higgs pair production as a relatively clean tool in such circumstances can be statistically limited and four top final states might well provide a more robust alternative route for discovery.
The aim of this work is to clarify both these avenues: building on sensitivity enhancements in discriminating SM four top production, we quantitatively analyse the sensitivity improvement when turning to motivated extensions of the SM.A further emphasis is given to the relevance of interference effects in four top final states.
This paper is organized as follows.In Sec.II, we provide details on event simulation and inclusive fiducial selection for the different (lepton-dominated) partonic collider processes that form the basis of this study: same-sign dilepton in association with b quark and light jets In the following, l ¼ e, μ and we include leptonic decays of τ leptons (see Refs. [1,3] for recent explorations by the ATLAS and CMS experiments).Subsequently, we detail in Sec.II our analysis implementation in terms of a graph neural network (GNN) that provides a particularly motivated approach to exploiting the various particle-level correlations expected in signal and background distributions.Section III is devoted to results and we first discuss the sensitivity yield for SM four top production (Sec.III A), which provides the baseline for the subsequent BSM sensitivity discussion in Secs.III B and III C. Our BSM discussion is split into addressing the relevance of GNN techniques when facing resonant and nonresonant BSM extensions to gauge the potential of the approach for different phenomenological situations.For resonant BSM extensions, we also touch on the relevance of signalbackground interference, which is known to be large in resonant ditop final states [15][16][17][18][19][20][21], but turns out to be negligible for the dominant four top production modes.A summary and conclusions are provided in Sec.IV.

A. Event simulation and fiducial selection
To prepare our datasets we need to preselect the events appropriate for 2SSDL and 3L final state topologies.Events are generated from simulating proton-proton collisions at ffiffi ffi s p ¼ 13 TeV using MADGRAPH_AMC@NLO [22] with leading order precision.These events are subsequently showered and hadronized using PYTHIA 8.3 [23].
To identify our inclusive search region we employ the following baseline selection criteria.Light leptons (electrons and muons) are defined from a threshold of p T > 10 GeV within the detector coverage of the electromagnetic calorimeter jηj < 2.5, where p T and η denote the transverse momentum and pseudorapidity, respectively.Selected b-tagged jets are identified from a threshold p T > 20 GeV within the tracker jηj < 2.5.Light flavor jets that fall into the hadronic calorimeter coverage, jηj < 4.5, are fed to the analysis if they satisfy p T > 20 GeV.Furthermore, we require a significant amount of missing transverse energy of at least 20 GeV.To isolate the mismeasurement of missing transverse energy ( ⃗ E T ) from jets we select the events with azimuthal angle difference jϕðjÞ − ϕðE T Þ !j > 0.2.Our analysis for the two signal topologies 2SSDL and 3L closely follows the ATLAS and CMS studies of Ref. [1][2][3][4][5][6] (see also [27][28][29] for precision SM analyses informing experimental studies).For the 2SSDL final state, we require at least 2 leptons, at least two b jets and at least one additional jet according to the above criteria.To select the events for the 3L search, we require at least three leptons, at least two b jets without restriction on the light flavor jets.Throughout, we employ a 70% flat b-tagging efficiency, which corresponds to a pessimistic b tagging working point [30] (see also [31]).In turn this removes large contributions from misidentified b jets [31] for the considered p T range of b jets, hence we neglect effects from misidentified b jets in this study.
Having defined our fiducial regions, we turn to the dominant SM backgrounds contributing to these final states: (1) The t tW AE channel with leptonic decays of the tops and W is the most dominant background for the 2SSDL case; however its contributions are subdominant for the 3L channel.(2) Conversely, background contributions from t tZ are significant for the 3L case, while contamination of the 2SSDL signal region from this channel is reduced (yet substantial) compared to others.
2b jets, with subsequent decays W → lν l , is the second large background for 2SSDL and remains a considerable background for the 3L selection.(4) W þ W − Z þ 2b jets with leptonic decays of vector bosons contributes as a background for both the channels, and yields the largest contamination in the 3L case.(5) ZW AE þ 2b jets is a subdominant background to both channels when both Z and W bosons decay leptonically.However, its contribution to 2SSDL is negligible and therefore dropped from Table I.The cross sections after decay and baseline selection are shown in Tables I and II contributing to 2SSDL and 3L channels, respectively.
We prepare datasets for the background class of our neural network setup by combining individual background samples according to their cross section to obtain a realistic composition expected at a fixed luminosity.Representative kinematic distributions are shown in Fig. 1, comparing signal and background at hadron level.It is evident that H T and M 4t 1 are the most discriminating observables and hardnessrelated observables can be expected as the dominant drivers of signal vs. background discrimination.The variables H T and M 4t are significantly correlated since both use energymomentum information of the process, as shown in Fig. 2. The neural network can exploit such correlations (also extending to other observables, see Fig. 3) in a nonrectangular and optimized way, thus enabling efficient signalbackground discrimination.We can therefore expect machine-learned observables to carry significant discriminating power in isolating (B)SM four top production.

B. Graph neural network architecture and data representation
Given the complex final state that can be expected from four top topologies, in addition to the process-specific correlations of final state objects, GNNs are ideal candidate architectures to discriminate the signal characteristics from the expected background.Applications of GNNs to particle physics data have a short yet successful history.Their versatility in efficiently discriminating BSM data from the SM expectation beyond traditional observables in a robust way has been highlighted in series of papers, ranging from designing anomaly detection methods [32][33][34], over jet tagging [31,35] to constraining EFT operator deformations in the BSM top sector fits [36].
Taking inspiration from this evolving success story, we employ a GNN to discriminate the four top signal from the relevant SM background in this work.To implement the graph structure, we use the DEEP GRAPH LIBRARY [37] with the PYTORCH [38] backend and choose an Edge Convolution (EdgeConv) network to classify signal and background (final state particle events are interfaced employing PYLHE [39]).
The GNN can be divided into (i) message passing followed by (ii) node readout.Edge convolution is known to be particularly suited for extracting internode information (edge) from the given low-level node features (i.e., particle-level properties, see below).The message passing function for the edge convolution is defined as Here, ⃗ x ðlÞ i represents the node features of node i in the lth message passing layer, with l ¼ 0 denoting the input node features of the graph.The neighborhood set N ðiÞ consists of all nodes in the graph connected to node i.The linear layers Θ and Φ take the input vectors ð⃗ x ðlÞ j − ⃗ x ðlÞ i Þ and ⃗ x ðlÞ i , respectively, and map them onto alternate dimension spaces where one applies an activation function on their vector sum.The dimensionality of the "hidden" feature space is chosen such that performance is optimized, while avoiding overfitting.
For the four top vs. background discrimination task, we use a fully connected bidirectional graph as input to the GNN.Each particle is represented by a node, which is associated with a node feature vector ½p T ; η; ϕ; E; m; PID (representing transverse momentum, pseudorapidity, azimuthal angle, energy, mass and particle identification number, respectively).For our analyses, after a fixed number of message passing steps Eq. (2.1), we employ a mean graph readout operation to the node features of the final message passing layer to extract a vector capturing the graph properties of each event.The graph representation obtained this way is then fed into a linear layer with a RELU activation function.The final linear layer maps the result to a two dimensional vector, normalized by the SoftMax function, which corresponds to the probability of an event arising from a signal or a background process.
We use the ADAM optimizer [40] to minimize the crossentropy loss function with an initial learning rate of 0.001.The network is trained separately with data corresponding TABLE II.Summary of background cross sections and Monte-Carlo uncertainties for the different SM processes contributing to the 3L selection, given our baseline cuts.The subscript on each particle denotes the decay channel as in Table I.

Processes
Cross Section (fb) Cross sections for the different SM processes having the most significant contributions to the 2SSDL background, in accordance with our baseline cuts.The subscript on each particle describes the particle's decay channel (l AE describes leptonic decay and h refers to a hadronic decay channel).We include the uncertainties from the Monte-Carlo simulation.
Processes Cross Section (fb) overtraining (we find that deeper networks lead to overtraining by investigating accuracy and loss of the network).
Some comments regarding similar analysis strategies for the four top final state as given above are in order: a variety of ML algorithms have been explored in Ref. [11] to assess their efficiency for the four top process, including the graph-based PARTICLENET [41] utilizing the EdgeConv operation. 2The authors of Ref. [11] demonstrate that PARTICLENET does indeed provide superior performance in comparison with the alternative algorithms that are considered there.We do find a qualitatively similar behavior.

A. Four tops at the LHC: Isolating the SM signal
The Standard Model four top final state has been explored by the ATLAS and CMS collaborations with considerable effort.ATLAS currently observes four tops with a sensitivity of 1.9σ over the background-only hypothesis with a production cross section of 26 fb [3].In comparison, CMS observes 2.6σ at a cross section of 12.6 fb [1].
To comment on the GNN performance for SM four top production, we first turn to the sensitivity estimates given by the backgrounds described in Sec.II.We first train our network for the 2SSDL samples for SM four top (t tt t) as the signal class and the corresponding backgrounds (listed in Table I) as the background class.A similar procedure is followed for the 3L final state.We demonstrate the network performance using the receiver operating characteristic (ROC) curves of the network which are shown in Fig. 5 for the two different final states under consideration.The areas under the curves (AUCs) of 0.994 for 2SSDL and 0.968 for 3L, imply a very high classifier performance, as expected by the correlations of Figs. 1 and 2 (as well as the findings of Ref. [11]).
With the GNN setup optimized, we can cast the sensitivity encoded in the ROC curve into exclusion limits by considering the significance , where N s is the number of signal events and N b is the number of background events at a particular luminosity.The number of signal event N s is given by σ s × ε sðGNNÞ × L and number of background event is given is given by N B is given by σ b × ε bðGNNÞ × L, where σ s=b is the cross section of signal and background after the baseline selection, L is the integrated luminosity and ε s=bðGNNÞ are the efficiencies obtained at the best working point from the ROC curve such that σ We also explore the impact of systematic uncertainties by considering 20% and 50% overall systematic uncertainties in the analysis.
For the 2SSDL final state, the four top production cross section can be measured as 4.4 fb with a systematic uncertainty of 50%, and 4.1 fb with a systematic uncertainty of 20%, at an integrated luminosity of 139 fb −1 with a significance of 2.7σ over the background-only hypothesis using the GNN.Although we cannot claim comparability with the realistic experimental environments, these findings indicate that the GNN selection could indeed provide an avenue to further hone the sensitivity in ATLAS and CMS by integrating these techniques into their analysis flow.For the 3L final state, the four top production cross section can be measured as 18.6 fb with a systematic uncertainty of 50%, and 17.3 fb with a systematic uncertainty of 20%, at an integrated luminosity of 139 fb −1 with a significance of 2 standard deviations over the background-only hypothesis using GNN, implying a similar potential improvement as for 3L.The performance of the GNN in either channel reveals excellent discriminating power for the busy multitop states at the LHC with prospects of improving cross section measurements significantly.

B. Nonresonant new interactions: Modified Higgs boson interactions
Having demonstrated that the GNN indeed provides a suitable approach for extracting the four top SM signal, we 2 PARTICLENET uses the k-nearest-neighbors algorithm for the embedding of particles to graphs according to their separation , while we rely on fully-connected graphs.
can now turn to the relevance of the GNN for departures of the four top final states from the SM expectation.New physics modifications of four top final states from nonresonant interactions have been considered in a range of phenomenological studies [8,9].To highlight the relevance of the above approach, we consider the so-called Ĥ parameter as a motivated EFT-related example; Ref. [10] particularly emphasized the relevance of four top final states for associated searches (bounds have since been provided by the CMS experiment in [1]).Considering this particular nonresonant modification is a well-motivated test bed to motivate a more comprehensive analysis of four top interactions from the perspective of effective field theory [9,36].The Ĥ parameter is the analog of the Ŵ; Ŷ operators [42] of the gauge sector and can be understood as an oblique correction taking the form where D μ is the covariant derivative action on the Higgs doublet Φ. 3 The presence of this interaction modifies the propagation of the physical Higgs boson (suppressing the Higgs width which is not relevant as the Higgs boson is probed off-shell in four top production) with associated coupling modifications of the Higgs boson's couplings to massive vectors V ¼ W AE ; Z and heavy fermions (here the top quark) for canonically normalized fields.The correlation of HVV interactions and H propagator resulting in a cancellation is a consequence of gauge symmetry [10] which highlights multi-top final states as particularly suitable process to put constraints on the interaction of Eq. (3.1) beyond Higgs coupling measurements.
To include the effects of Ĥ in our study, we modify the HELAS routines [47,48] of our pp → t tt t implementation to reflect the modifications of propagators, vertices and their cancellation to linear order in the amplitude.As a benchmark for nonresonant new interactions, we change Ĥ ¼ 0.1 as a new physics modification (this value is motivated from the HL-LHC measurement sensitivity provided in Ref. [10]).We select a SM-rich sample by further applying a cut on H T greater than 900 GeV to enable a coupling measurement that is more geared to SM modifications, see Fig. 1.We then train a GNN network to discriminate between SM interaction with Ĥ ¼ 0 from the nonresonant new interaction with Ĥ ¼ 0.1.The kinematical differences are not dramatic (which is highlighted by the ∼10% sensitivity in the first place), and therefore further motivate the inclusion of as much correlated information as possible achievable through GNN applications.
We find an AUC of 60%, which is testimony to the difficulty of extracting electroweak properties from QCDbusy final states.The observed sensitivity can be used to set limits for in the parameter range Ĥ ∈ ½0.0; 0.10 shown in Fig. 6.The blue curve shows the required luminosity to obtain the 3σ bound on that particular Ĥ value, whereas the red curve shows the significance obtained at a fixed luminosity L ¼ 1000 fb −1 .Our GNN analysis again gives rise to an expected sensitivity improvement beyond the estimates of cut-and-count analyses [1,10].While the realistic analysis in the LHC environment is likely more limited in sensitivity, the addition of the GNN techniques described above could significantly improve the sensitivity to Ĥ considerably, which also indicates an improved sensitivity to generic EFT deformations.

C. Resonant new interactions: Interference effects and 2HDM reach
Four top final states in the context of resonant extensions of the SM have been studied in a range of analyses [49][50][51][52][53][54][55][56][57], typically relying on traditional collider observables alongside cut-and-count strategies.While new resonant structures lend themselves to such approaches, we can also FIG. 6. Significance as a function of Ĥ at 13 TeV LHC with a 1000 fb −1 integrated luminosity.We also present the required luminosity as a 3σ confidence level for different Ĥ values. 3It is worthwhile mentioning that the modification of the propagator probes a genuine direction of Higgs effective field theory (HEFT), that can be radiatively sourced in the so-called κ framework [43], see e.g.[44,45].The associated coupling modifications can then be understood as a probe of the linearity of the Higgs boson's interactions in the SM and its dimension-6 extension [46].expect a significant sensitivity enhancement when turning to ML approaches (see also [58][59][60][61][62][63]).
Before we comment on the GNN performance, however, we will perform a qualitative analysis of the interference effects that are known to be large in direct gluon fusion production gg → t t of extra scalar resonances [13,14].Such interference effects can severely limit the sensitivity estimates and are not straightforward to include in ML-based selections due to negative weights.To this end we introduce a simplified scalar resonance S with Higgs-like couplings to the top quark 4 keeping the width of S as a free parameter in a scan to qualitatively assess S resonance-distortion (ξ S plays the role of the Higgs coupling modifier for S ¼ H).To study the interference effects in the four top final states, we compare the resonance structure stemming for the new-physics-only (resonance) interactions dσ new ∼ jM res j 2 dLIPS ð3:6Þ with interference-only contributions The invariant mass distribution of m tt system closest to the mass of the new scalar S as detailed in the text. 4We model the phenomenological Lagrangian with FEYNRULES [64,65] and export the relevant Feynman Rules in the UFO [66] format.We use the same toolchain as before with the addition of MADSPIN [67,68].
where "bkg" refers to any nonresonant amplitude contribution (e.g., the continuum QCD background).A good qualitative understanding of the signal distribution distortion can be obtained by isolating the peaked m tt distribution: we construct all possible combinations of m tt and choose the invariant mass which is closest to the candidate mass M S .The resulting distributions are shown in Fig. 7 for a range of mass M S ¼ ½500; 600; 700; 800; 1000 GeV and width choices Γ S =M S ¼ 5%; 10%; 20% for ξ S ¼ 1.
From the plots shown in Fig. 7, we can infer that the interference effects that distort the mass peak are comparably small.Closer to the decoupling limit ξ S → 0 we will encounter more sensitivity-limiting distortion, however this will happen at quickly vanishing BSM resonance cross sections so that a (ML-assisted) bump hunt will not provide any sensitivity when considering the backgrounds.In the case of imaginary phases of ξ S , i.e., CP-odd coupling structures, the interference effects will be qualitatively similar see Refs.[15] (they can also vanish in processes when interference with CP-even backgrounds removes such effects [60]).
Having identified interference effects as largely irrelevant, we return to the GNN analysis of the BSM resonance structures.We train our network for the same two final states, this time with the added new interactions (i.e., the newly added CP-even scalar with ξ S ¼ 1) for the different benchmark points listed in Table III.The ROC curves for one such benchmark point (M S ¼ 600 GeV) for both the final states and their corresponding AUCs are shown in Fig. 8.The corresponding significance for each benchmark point for the different decay channels is also listed in Table III for ξ S ¼ 1 and Γ S =M S ¼ 0.1.We interpret these limits as bounds on the coupling modifier ξ S as shown in Fig. 9.
Although we can indeed put constraints ξ S < 1 (predominantly through the 2SSDL channel), the resulting bounds are relatively weak.For instance in singlet mixing scenarios, the parameter range that this analysis is sensitive to is already highly constrained from Higgs signal strength measurements.Turning to phenomenologically richer scenarios we further examine the 2HDM of type II, which features new neutral CP odd and CP even states.In this model, the Higgs sector consists of two distinct complex doublets, both with a nonzero vacuum expectation value (vev) (reviews can be found in Refs.[69,70]).One of these doublets couples to the up-type quarks and the other to down-type quarks and leptons in the type II model, thereby leading to fermion mass terms.After electroweak symmetry breaking, one is left with 5 Higgs bosons; 4 of which are CP even; h, H and H AE , and a CP odd particle, A. The lighter of the two neutral CP even states is identified as the observed Higgs boson.The phenomenology of this model then depends on the masses of the new Higgs bosons, the ratio of the two vevs, defined as tan β ¼ v 2 =v 1 and cosðβ − αÞ, where α is the mixing angle between the two neutral CP even states.The simplified Lagrangian of Eq. (3.5) can be mapped onto the 2HDM type II Yukawa sector The 2HDM is a particularly well motivated BSM model, as it is, in principle, capable of resolving a number of tensions between current experimental values and SM predictions, as well as opening avenues to satisfy the Sakharov criteria for required baryogenesis [71].As such, this model is well studied in the literature [72][73][74][75][76][77][78][79][80].Here we use this model to explore the validity of the benchmark points examined above.In order to satisfy the theoretical considerations the masses of the new particles are set to be entirely degenerate.
Similarly, to fall in line with the signal strength data of the SM Higgs we set cosðβ − αÞ ¼ 0; the alignment limit which recovers exactly the SM phenomenology of h.
To examine the mass range of the above benchmark points in the face of the collider searches for additional Higgs bosons we make use of the 2HDECAY [81][82][83][84][85][86] and HIGGSBOUNDS [87][88][89][90][91][92][93] packages.Additionally, we extrapolate the LHC search data to 3000 fb −1 at 13 TeV.Performing a scan over 20000 random points, the results of this analysis are shown in Fig. 10, with points in blue allowed by current and extrapolated data, while orange points are allowed only by the current data.The improvement in sensitivity is lead by searches for the decays H þ → t b and H 0 → τ þ τ − [94,95].This shows that a large swathe of the parameter space that we take for benchmark points in the above analysis is possible within the framework of the 2HDM of type II.
Given that four top final states are far less abundant than other processes, they can be expected to exhibit a reduced sensitivity to new states (this is already visible from Fig. 9).Nonetheless, repeating the above analysis also including the CP odd state, we find that complementary sensitivity can be achieved for the 2HDM scenario with the GNNassisted resonance search detailed above: By overlaying the constraints found from 2SSDL final states (shaded area) we see that there is a nontrivial parameter range of the 2HDM type II's alignment limit that can be accessed four top final states at 2σ confidence and 3000 fb −1 .This example contains only the more sensitive channel of the ones discussed in this work and additional final states are likely to further improve this result.

IV. SUMMARY AND CONCLUSIONS
The growing sensitivity to four top final states at the Large Hadron Collider [1][2][3][4][5][6] bears great potential for the discovery of new physics beyond the Standard Model.While the complexity of the final states requires a dedicated search strategy compared to lower multiplicity processes, the plethora of kinematic information that can be accessed in these processes to highlight departures from the SM can be formidably exploited by means of adapted search strategies, e.g.building on machine learning approaches.These do not only provide support in establishing sensitivity to SM four top production as shown in Sec.III A (see also [11]), but they become particularly relevant when four top final states will be scrutinized from a BSM perspective in the future.We have demonstrated that significant sensitivity to wellmotivated representative nonresonant physics can be achieved (Sec.III B), but also that sensitivity to resonances is gained by turning to graph neural networks (Sec.III C).The GNN exploits the hierarchical graph structure of particle physics events that is particularly accentuated when dealing with new resonances.In this case, the four top final states can also fill a potential sensitivity gap in s-channel gluon fusion production of top-philic states gg → t t that results from destructive signal-background interference, distorting or removing the FIG.10.Scan of the parameter space of the type II 2HDM in the face of current collider data, and LHC search data extrapolated to 3000 fb −1 at 13 TeV, with a 2σ confidence level exclusion contour from 2SSDL final states.Points in blue are allowed by current and expected future datasets, orange points by current data only.The shaded area gives the sensitivity of the 2SSDL resonance search detailed in the text at 3000 fb −1 .Interference effects from gg → t t sensitivity estimates are not included.FIG. 9. 2σ confidence level limits on scalar mass with respect to CP-even and CP-odd couplings for 3000 fb −1 for 13 TeV LHC collisions.
resonance structure.The dominant partonic channels of four top final states show no such behavior and therefore remain robust processes even when interference removes sensitivity in top pair gluon fusion production.In particular, this can add complementary sensitivity to BSM searches for (but not limited to) two Higgs doublet models as we have shown in a representative parameter scan for the 2HDM type II.
b quarks ð2SSDLÞ; ð1:1aÞ and three lepton production in association with b and light flavor jets

Þb b 12 .
FIG. 1. Normalized distributions of the kinematic variables related to 2SSDL channel for SM four top signal and SM backgrounds contributing to the process at the LHC running at 13 TeV.

FIG. 5 .
FIG. 5. ROC curves for discriminating the SM four top quark signal in the (a) 2SSDL and (b) 3L final states.

TABLE III .
Significances for different masses M S for a coupling choice ξ S ¼ 1 and Γ S =M S ¼ 0.1, Eq. (3.5).The luminosity is taken to be 3000 fb −1 for 13 TeV LHC collisions.
FIG. 8. ROC curves for training a four top quark signal for a BSM CP-even scalar with M S ¼ 600 GeV for (a) 2SSDL and (b) 3L final states.