Direct optimisation of the discovery significance when training neural networks to search for new physics in particle colliders

TL;DR

This paper proposes two new loss functions for neural networks that directly optimize the statistical significance in particle physics searches, potentially improving detection sensitivity in collider experiments.

Contribution

Introduction of two novel loss functions that directly maximize significance estimates for neural network training in physics searches, outperforming traditional methods under systematic uncertainties.

Findings

Loss functions based on $s/\sqrt{s+b}$ and $Z_A$ improve search sensitivity.

The $Z_A$-based loss outperforms binary cross entropy in systematic uncertainty scenarios.

Application demonstrated in a toy SUSY search with LHC data.

Abstract

We introduce two new loss functions designed to directly optimise the statistical significance of the expected number of signal events when training neural networks to classify events as signal or background in the scenario of a search for new physics at a particle collider. The loss functions are designed to directly maximise commonly used estimates of the statistical significance, $s/\sqrt{s+b}$, and the Asimov estimate, $Z_A$. We consider their use in a toy SUSY search with 30~fb$^{-1}$ of 14~TeV data collected at the LHC. In the case that the search for the SUSY model is dominated by systematic uncertainties, it is found that the loss function based on $Z_A$ can outperform the binary cross entropy in defining an optimal search region.