Kinematic Reconstruction of Top-Quark Pairs Using Analytical and Machine-Learning Methods with CMS Open Data
summary
The gist
The reconstruction of the top-quark-pair (t t¯) kinematics in the dilepton decay channel is challenging due to the presence of two undetected neutrinos This work develops improved kinematic…
In short
This work improves kinematic reconstruction of top-quark pairs (t t¯) from dilepton decays, which is hard because two neutrinos are invisible. The study compares existing methods like Loose Kinematic Reconstruction with machine learning approaches (MLP and Flow). ML methods significantly enhance the resolution of the t t¯ invariant mass, showing that they can learn physics corrections beyond simple approximations.
Key concepts
- Loose Kinematic Reconstruction (LKR)
- An existing technique that reconstructs only the t t¯ system as a whole by exploiting kinematic similarities between dilepton and neutrino-pair systems. It is used as a baseline method for comparison, but it has limitations in resolving the invariant mass.
- Machine Learning Approaches
- Two ML methods, a residual multilayer perceptron (MLP) and a conditional normalizing flow, were developed to improve reconstruction. These methods learned genuine physics-driven corrections related to the neutrino-pair approximation that analytical algorithms miss.
- Invariant Mass M(t t¯)
- This is the key quantity being reconstructed, representing the total mass of the top-quark pair. Accurately reconstructing this mass is crucial for measuring the top quark's mass (m t), and ML methods provide a substantial improvement in its resolution compared to traditional methods.
Terminology used across episodes
This episode discusses
- Kinematic Reconstruction of Top-Quark Pairs Using Analytical and Machine-Learning Methods with CMS Open Data · Paper Radio
- What is the Top Quark Mass?
- Higgs mass and vacuum stability in the Standard Model at NNLO
- The top quark and Higgs boson masses and the stability of the electroweak vacuum
- Spin Correlation Effects in Top Quark Pair Production at the LHC
- Review of top quark mass measurements in CMS
- Measurement of the t t bar cross section and its ratio to the Z production cross section using pp collisions at sqrt s = 13.6 TeV with the ATLAS detector
- Measurement of the t production cross section, the top quark mass, and the strong coupling constant using dilepton events in pp collisions at sqrt s = 13 TeV
- Measurement of differential top-quark pair production cross sections in pp collisions at sqrt(s) = 7 TeV
- Measurement of the t-tbar production cross section in the e-mu channel in proton-proton collisions at sqrt(s) = 7 and 8 TeV
- Algebraic approach to solve t t bar dilepton equations
- Analytical solution of ttbar dilepton equations
- Measurement of double-differential cross sections for top quark pair production in pp collisions at sqrt(s) = 8 TeV and impact on parton distribution functions
- Measurement of t t normalised multi-differential cross sections in pp collisions at sqrt s =13 TeV, and simultaneous determination of the strong coupling strength, top quark pole mass, and parton distribution functions
- Differential cross section measurements for the production of top quark pairs and of additional jets using dilepton events from pp collisions at sqrt s = 13 TeV
- Measurement of the top quark pole mass using t t bar +jet events in the dilepton final state in proton-proton collisions at sqrt s = 13 TeV
- The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations
- PYTHIA 6.4 Physics and Manual
- The anti-k t jet clustering algorithm
- Layer Normalization
- Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning
The paper
Kinematic Reconstruction of Top-Quark Pairs Using Analytical and Machine-Learning Methods with CMS Open Data · Read on arXiv
V. Aushev, M. Spitsyn, R. Yermolenko, O. Zenaiev
Taras Shevchenko National University of Kyiv · Institute for Nuclear Research, National Academy of Sciences of Ukraine · II Institute for Theoretical Physics, Hamburg University
Transcript
Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.
Kai: We're looking at the paper: "Kinematic Reconstruction of Top-Quark Pairs Using Analytical and Machine-Learning Methods with CMS Open Data".
Mira: Here's the gist of the paper. The reconstruction of the top-quark-pair (t t¯) kinematics in the dilepton decay channel is challenging due to the presence of two undetected neutrinos This work develops improved kinematic reconstruction methods, including machine learning approaches, to enhance the resolution of the reconstructed t t¯ invariant mass Kinematic Reconstruction Baseline and Improvements Existing kinematic reconstruction techniques for dilepton t t¯ events include full and loose kinematic reconstruction (FKR) The FKR recovers the three-momenta of the two neutrinos using missing transverse energy (MET) and known masses, but it requires fixing the m t value a priori An alternative loose kinematic reconstruction (LKR) reconstructs only the t t¯ system as a whole, exploiting kinematic similarities between the dilepton and neutrino-pair systems The LKRv3 modification exploits the approximate equality of invariant masses and rapidities of the dilepton and neutrino-pair systems Machine Learning Approaches Two machine learning (ML) based approaches, a residual multilayer perceptron and a conditional normalizing flow, were developed to improve reconstruction Relative to LKRv3, these ML approaches improve the resolution of the invariant mass by approximately 30%, transverse momentum by 20%, and rapidity by 15% The MLP and Flow methods demonstrate similar performance when trained on the same input information Performance Metrics at Detector Level The performance of the methods is evaluated in terms of efficiency, bias, and resolution Table 1 shows that the MLP and Flow improve sigma(M(t t¯)) by 35% with respect to LKR and by 30% with respect to LKRv3 The improvement is present in all bins of M(t t¯), pT (t t¯) and y(t t¯) Generator Level Analysis The networks are also trained and applied at generator level, where the inputs are the true leptons, b quarks, and neutrino pair transverse momentum without CSV discriminants The networks reduce sigma(M(t t¯)) to 50 GeV, a factor of 1.8, indicating that most of the limitation of analytical algorithms comes from the neutrino approximation The results demonstrate that ML methods can substantially improve the performance of analytical kinematic reconstruction, particularly for M(t t¯), which is most relevant for measurements of m t The transfer test demonstrates that the neural network learns a genuine physics-driven correction to the neutrino-pair approximation The achievable resolution is limited primarily by the information available in the reconstructed event and the intrinsic limitations of the neutrino-pair approximation, rather than by the capacity of the neural network Pruning redundant reasoning tokens The reconstruction of top-quark-pair (t t¯) kinematics in the dilepton decay channel is challenging due to the presence of two undetected neutrinos The study tests existing kinematic reconstruction methods using a Monte Carlo sample with t t¯ events from CMS Open Data at a centre-of-mass energy of 7 TeV How the training signal is built The MLP is trained with the L1 loss (mean absolute error), as it estimates the median of the targets for given inputs and gives a smaller bias and a better conditioned response than the L2 loss The Flow is trained by minimizing the negative log-likelihood of the true values Where the gains come from The improvement in resolution is largest near the threshold, where the best achieved M(t t¯) resolution is 30 GeV compared to 78 GeV for LKR The generator-level results show that a substantial part of the uncertainty in t t¯ kinematics originates from the intrinsic limitations of the neutrino approximation, rather than from detector resolution and/or b-jet assignment The transfer test demonstrates that the neural network learns a genuine kinematic correction The kinematic reconstruction of t t¯ events in the dilepton decay channel was studied using the detector-simulated CMS MC sample available through the CMS Open Data The MLP and Flow achieve comparable reconstruction performance with the same input information The MLP is approximately an order of magnitude less computationally expensive and provides a better-conditioned response matrix These results demonstrate that ML methods can substantially improve the performance of analytical kinematic reconstruction, particularly for M(t t¯), which is most relevant for measurements of m t The achievable resolution is limited primarily by the information available in the reconstructed event and the intrinsic limitations of the neutrino-pair approximation, rather than by the capacity of the neural network Pruning redundant reasoning tokens The top quark is one of the heaviest known elementary particle
1, 2: due to its large mass The study of the top quark is one of the key areas in the physics program at the Large Hadron Collider (LHC) where top quarks are produced copiously in p p collisions, predominantly as pairs of top quark and antiquark (t t¯) via the strong interaction The top quark has a very short lifetime of approximately 5 × 10−25 s and thus decays via the weak force before hadronization (which is driven by non-perturbative strong interactions and would happen at the time scale of ∼ 10−23s) How the training signal is built The inputs and targets are standardized to zero mean and unit variance, with the means and variances computed on the training events and stored together with the network The MLP is trained with the L1 loss (mean absolute error), as it estimates the median of the targets for given inputs Where the gains come from The two modifications of the LKR were developed, with LKRv2 applying more stringent requirements to jet-lepton combinations and LKRv3 using approximate equality of invariant masses and rapidities The ML methods learn a genuine physics-driven correction to the neutrino-pair approximation, rather than merely exploiting detector-specific effects The resolution of M(t t¯) is dominated by the neutrino pair approximation, while the resolution of pT (t t¯) is dominated by finite detector resolution The kinematic reconstruction of t t¯ events in the dilepton decay channel was studied using the detector-simulated CMS MC sample available through the CMS Open Data The performance of existing reconstruction methods was evaluated in terms of efficiency, resolution and bias As a baseline for further development, the LKR method was chosen which provides a high reconstruction efficiency of about 98% and avoids the dependence of the FKR on a fixed value of m t The results demonstrate that ML methods can substantially improve the performance of analytical kinematic reconstruction, particularly for M(t t¯), which is most relevant for measurements of m t The achievable resolution is limited primarily by the information available in the reconstructed event and the intrinsic limitations of the neutrino-pair approximation, rather than by the capacity of the neural network Pruning redundant reasoning tokens The final state of an event with t t¯ can be classified by further decays of W bosons into three final states: (1) dilepton decay channel (10.5%), (2) semileptonic decay channel (43.8%), and (3) fully hadronic decay channel (45.7%) Electrons and muons are easily reconstructed with very good resolution, while tau leptons are more difficult to reconstruct experimentally How the training signal is built The inputs of the neural networks are based on 24 variables: the four-momenta of l−, l+, and of the two jets selected by LKRv3 and the missing transverse momentum (18 variables, “raw”), additional four lepton–jet masses M(l± j1,2) and two CSV b-tagging discriminants of the two jets The selection of input jets is done in the same way as in the LKRv2/LKRv3, therefore the NN methods have the same reconstruction efficiency of 96.5% Where the gains come from The importance of M(l j) masses (30% for M(t t¯)) and their small gain in Table 3 are not contradictory, as after shuffling, they are inconsistent with the four-momenta The MET is important mainly for pT (t t¯), in agreement with the generator-level study The CSV discriminants change the overall resolution by only 7%, since their effect is concentrated near the threshold Pruning redundant reasoning tokens The top quark has a very short lifetime of approximately 5 × 10−25 s and thus decays via the weak force before hadronization (which is driven by non-perturbative strong interactions and would happen at the time scale of ∼ 10−23s) In the dilepton t t¯ decay channel kinematic reconstruction is especially challenging due to the presence of two undetected neutrinos Where the gains come from The deviation of true neutrino pair kinematics from the LKRv3 approximation correlates with observed decay products, which a neural network can learn The networks are trained on test events only, and the same events are used for all comparisons The augmentation of training events by random rotation around the beam axis increases the diversity of the training sample without altering regression targets Pruning redundant reasoning tokens The top quark has a very short lifetime of approximately 5 × 10−25 s and thus decays via the weak force before hadronization (which is driven by non-perturbative strong interactions and would happen at the time scale of ∼ 10−23s) The final state of an event with t t¯ can be classified by further decays of W bosons into three final states: (1) dilepton decay channel (10.5%), (2) semileptonic decay channel (43.8%), and (3) fully hadronic decay channel (45.7%) Where the gains come from The deviation of true neutrino pair kinematics from the LKRv3 approximation correlates with observed decay products, which a neural network can learn The augmentation of training events by random rotation around the beam axis increases the diversity of the training sample without altering regression targets Pruning redundant reasoning tokens The top quark has a very short lifetime of approximately 5 × 10−25 s and thus decays via the weak force before hadronization (which is driven by non-perturbative strong interactions and would happen at the time scale of ∼ 10−23s) The final state of an event with t t¯ can be classified by further decays of W bosons into three final states: (1) dilepton decay channel (10.5%), (2) semileptonic decay channel (43.8%), and (3) fully hadronic decay channel (45.
Paper summary: Kai: So to wrap up this paper, the work by Aushev et al. on "Kinematic Reconstruction of Top-Quark Pairs Using Analytical and Machine-Learning Methods with CMS Open Data" shows how machine learning can substantially improve the performance of analytical kinematic reconstruction methods.
Mira: They developed these neural networks, specifically a residual multilayer perceptron and a conditional normalizing flow, to tackle the challenge of reconstructing top quark pair kinematics in the dilepton decay channel where two neutrinos are undetected.
Lev: The key findings are that these ML approaches improve the resolution of the invariant mass by about thirty percent relative to LKRv3 at detector level, and they show similar performance when trained on identical input information.
Kai: And they also achieved a factor of one point eight reduction in the uncertainty for the top quark pair invariant mass when training at generator level, which suggests that most of the analytical algorithm's limitations come from that neutrino approximation.
Mira: The authors conclude that while these neural networks learn a genuine physics driven correction to the neutrino pair approximation, they also state that the achievable resolution is limited primarily by the information available in the reconstructed event and those intrinsic limitations of the neutrino pair approximation.
Lev: So, what this means for us is that we can't just rely on an arbitrarily powerful network to fix every issue; we still have to improve our input reconstruction and understand those missing neutrinos better.
Kai: It shows that ML methods are a useful tool here, especially for M(t t bar) measurements, but the fundamental physics limitation remains what the authors point out.
Conclusion: Kai: So we're wrapping up this talk on "Kinematic Reconstruction of Top-Quark Pairs Using Analytical and Machine Learning Methods with CMS Open Data." Basically, they took those tricky top quark pair events in the dilepton channel where you have two neutrinos hiding, and they used new machine learning tools to get better at figuring out the kinematics.
Mira: Right. They looked at existing methods like loose kinematic reconstruction, LKRv3, and then introduced these neural networks—a multilayer perceptron and a conditional normalizing flow—to sharpen the measurement of the top quark pair invariant mass.
Lev: From an engineering standpoint, what’s really interesting is that the machine learning models learned a correction to how those two neutrinos behave in this approximation. It’s not just fitting noise; it's learning some underlying physics correction.
Kai: That's what they called a genuine physics-driven correction, which is important because it means the network isn't just memorizing detector quirks; it’s finding something new about the neutrino pair kinematics itself.
Mira: But they also stressed that their ultimate success depends on two things: how much information you actually have in your reconstructed event, and those fundamental limitations of treating the neutrinos as a simple approximation.
Lev: So, while the network helps boost the resolution by thirty percent at detector level, the real bottleneck is still what you can see from the raw data.
Kai: Exactly. The paper shows that for measuring something like top quark mass, this ML help is significant, but it doesn't magically fix all our problems with missing information or approximations.
Mira: It suggests that machine learning isn't a magic solution for everything; it’s a powerful tool to see the corrections where they actually live.
Lev: It means we need to keep improving how we reconstruct the initial event data if we want better results than what these networks can provide on their own.
More episodes
- Daily Summary for 2026-11-19
- 2512.21290-Model-independent ZH production cross section at FCC-ee
- 2610.21982-First Observation of the Radiative Decay D 0 to gamma omega
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- Daily Summary for 2026-11-18
- 2512.08065-Searches for Light Dark Matter and Evidence of Coherent Elastic Neutrino-Nucleus Scattering of Solar Neutrinos with the LUX-ZEPLIN (LZ) Experiment
- Daily Summary for 2026-11-18
- 2508.18022-Precision Measurements of the Electroweak Mixing Angle in the Region of the Z pole at the Large Hadron Collider
- 2512.23992-Prospect for measurement of CP-violating parameters of B s 0 to phi gamma at the Tera Z factory