$\nu$-flows: Conditional neutrino regression

Abstract

We present \nuν-Flows, a novel method for restricting the likelihood space of neutrino kinematics in high-energy collider experiments using conditional normalising flows and deep invertible neural networks. This method allows the recovery of the full neutrino momentum which is usually left as a free parameter and permits one to sample neutrino values under a learned conditional likelihood given event observations. We demonstrate the success of \nuν-Flows in a case study by applying it to simulated semileptonic t\bar{t}tt‾ events and show that it can lead to more accurate momentum reconstruction, particularly of the longitudinal coordinate. We also show that this has direct benefits in a downstream task of jet association, leading to an improvement of up to a factor of 1.41 compared to conventional methods.

CaloFlow for CaloChallenge dataset 1

“…In fact, CALOFLOW proved to be the first ever generative model in HEP that could pass the following stringent test: a binary classifier trained on the raw voxels of (CALOFLOW) generated vs. (GEANT4) reference showers could not distinguish the two with 100% accuracy. Normalizing Flows showed similar good performance in other generative tasks in high-energy physics [21][22][23][24][25][26][27][28][29][30][31][32][33][34].…”

MadNIS - Neural multi-channel importance sampling

“…This rough online training is successful because normalizing flows, or invertible networks (INNs) [19,20], are especially well-suited, stable, and precise in LHC physics applications [21]. This has been shown in many instances, including event generation [22][23][24][25], detector simulations [26][27][28][29], unfolding or inverse simulations [20,30], kinematic reconstruction [31], Bayesian inference [32,33], or inference using the matrix element method [34]. On the other hand, for expensive integrands online training is clearly not optimal, because it does not make use of all previously generated data at subsequent stages of the network training.…”

Two invertible networks for the matrix element method

“…Once we control the forward direction with NN-based event generators [51][52][53][54][55][56], conditional GANs and INNs also allow us to invert the simulation chain, to unfold detector effects [57][58][59] or to extract the hard scattering process in a statistically consistent manner [60,61]. The fully calibrated inverted simulation uses the same conditional INN (cINN) as simulation-based inference [62,63] or kinematic reconstruction [64]. Obviously, any application of (generative) networks to LHC physics requires an uncertainty treatment [56,65].…”