Punzi-loss:

F Abudinén ,S Lacaprara ,J Kahn ,L Corona ,P Urquijo ,P Feichtinger ,M Bertemes ,Y Sato ,S.-H Park ,P Goldenzweig ,M Hohmann ,F Metzner ,G Inguglia ,T Humair ,H Haigh ,K Lieret ,M De Nuccio ,T Kuhr ,U Tamponi ,R Žlebčík ,G De Pietro ,Thomas M Keck ,F Meier ,C Schwanda ,M Milesi ,R Maiti ,I Komarov ,J Gemmler ,M Campajola ,C Pulvermacher ,S Cunliffe ,J.-F Krohn ,M Eliachevitch ,W Sutcliffe ,S Bilokin ,A Martini ,M Ritter ,G Casarosa ,M T Prim ,L Zani ,A Gottmann ,E Graziani ,F Tenchini ,T Ferber ,S Dey ,A Zupanc

doi:10.1140/epjc/s10052-022-10070-0

Abstract

We present the novel implementation of a non-differentiable metric approximation and a corresponding loss-scheduling aimed at the search for new particles of unknown mass in high energy physics experiments. We call the loss-scheduling, based on the minimisation of a figure-of-merit related function typical of particle physics, a Punzi-loss function, and the neural network that utilises this loss function a Punzi-net. We show that the Punzi-net outperforms standard multivariate analysis techniques and generalises well to mass hypotheses for which it was not trained. This is achieved by training a single classifier that provides a coherent and optimal classification of all signal hypotheses over the whole search space. Our result constitutes a complementary approach to fully differentiable analyses in particle physics. We implemented this work using PyTorch and provide users full access to a public repository containing all the codes and a training example.

Highlights

The standard model (SM) of particle physics is the theoretical framework that describes fundamental interactions and the fundamental constituents of matter
At the Belle II experiment, this search was performed with the commissioning data for the specific case of invisible decays of the Z boson [10], a final state in which only the two muons produced by the electron-positron annihilation can be reconstructed
We have demonstrated that it is possible to implement a non-differentiable metric approximation and a corresponding loss-scheduling, combining the approach of particle physics and that of machine learning

Summary

Introduction

The standard model (SM) of particle physics is the theoretical framework that describes fundamental interactions and the fundamental constituents of matter. When searching for new particles, for example, in a collider experiment, one of the main challenges is correctly reconstructing and identifying the new particles (the signal) and rejecting any (or most) contributions from potential background sources. This is a common problem referred to as event classification. Typical MVA methods in use in the field of particle physics include, but are not limited to, decision trees, boosted decision trees (BDTs) [1], or shallow and deep neural networks (NNs) [2]. As a benchmark study to test the performance of the Punzi-loss and compare it to other techniques, we consider the search for invisible decays of the hypothetical Z boson produced in the reaction e+e− → μ+μ− Z at the Belle II experiment [4,5] at the SuperKEKB collider [6], based on MC simulations

Neural networks

Figure of merit

Punzi-loss

Training strategy

Results

Conclusions