Two invisible neutrinos no longer force an analytic guess: machine learning cuts top-pair mass resolution by another 30%

Dilepton top-pair reconstruction gains a machine-learning route: networks regress the top-pair four-momenta directly from visible decay products instead of relying only on an analytic neutrino approximation. The earlier loose kinematic reconstruction, LKR, reconstructed only the top-pair system as a whole and set the longitudinal momentum and energy of the neutrino pair equal to those of the dilepton system, so some events produced unphysical values that had to be truncated. On the 7 TeV dilepton simulated sample from CMS Open Data, relative to the improved LKRv3, the multilayer perceptron and the normalizing flow improve the resolution of the invariant mass, transverse momentum and rapidity by about 30%, 20% and 15%, respectively.

The deviation the networks learn correlates with the kinematics of visible particles, so the three top-pair quantities can be predicted directly from leptons, jets and missing transverse momentum. The multilayer perceptron stacks residual blocks of hidden layers, while the normalizing flow uses autoregressive spline transformations to give a per-event probability density; both take the same 24 reconstructed inputs. On about test events, the multilayer perceptron lowers the invariant-mass resolution from 107.7 GeV for LKRv3 to 75.2 GeV, and the flow to 74.1 GeV, differences far larger than the statistical uncertainty and the spread across five training seeds.

A generator-level comparison separates the neutrino approximation from detector effects, showing that the networks correct the physics approximation itself. At generator level, where only the neutrinos are unknown and no detector resolution applies, LKRv3 still leaves 90.1 GeV invariant-mass resolution, while the multilayer perceptron drops to 50.3 GeV. In a transfer test that trains at generator level and applies the network to detector-level inputs, the networks still beat LKRv3 but retain only part of the improvement from detector-level training.

The results rest on the 7 TeV TTJets simulated sample in CMS Open Data, use no real collision data, and cover only the dilepton decay channel; the authors propose next testing on samples containing non-top-pair events to assess background rejection and signal significance, and checking robustness at different centre-of-mass energies and eventually on ATLAS open-data simulations.

The authors state that the achievable resolution is limited mainly by the neutrino-pair approximation and the information available in the reconstructed event, not by network capacity; the network estimates carry a bias toward the densely populated region of the spectrum, whose dependence on the simulated spectrum shape must be treated as a systematic uncertainty; precise top-mass extraction near the threshold also requires handling toponium, which would need the invariant-mass resolution in that region improved to the GeV level.