Transferability of Deep Learning Models in Searches for New Physics at Colliders

TL;DR

This study evaluates how well deep learning models trained on specific new physics signals at colliders can transfer their detection capabilities to other signals, demonstrating their robustness and potential for broad applicability.

Contribution

It provides a systematic assessment of the transferability of deep neural networks across different beyond standard model signals in collider searches.

Findings

Deep neural networks maintain similar limits across different vector-like T-quark signals.

Networks trained on flavor changing neutral current signals are less transferable but still effective.

Deep learning models can detect new physics signals even if not explicitly trained on them.

Abstract

In this work we assess the transferability of deep learning models to detect beyond the standard model signals. For this we trained Deep Neural Networks on three different signal models: $tZ$ production via a flavour changing neutral current, pair-production of vector-like $T$-quarks via standard model gluon fusion and via a heavy gluon decay in a grid of 3 mass points: 1, 1.2 and 1.4 TeV. These networks were trained with $t\bar{t}$, $Z$+jets and dibosons as the main backgrounds. Limits were derived for each signal benchmark using the inference of networks trained on each signal independently, so that we can quantify the degradation of their discriminative power across different signal processes. We determine that the limits are compatible within uncertainties for all networks trained on signals with vector-like $T$-quarks, whether they are produced via heavy gluon decay or standard model gluon fusion. The network trained on flavour changing neutral current signal, while struggling the most on the other signals, still produce reasonable limits. These results indicate that deep learning models are capable of providing sensitivity in the search for new physics even if it manifests itself in models not assumed during training.