Robust and Generalizable Background Subtraction on Images of Calorimeter Jets using Unsupervised Generative Learning

TL;DR

This work represents the first use of unsupervised unpaired generative models for full detector jet background subtraction and offers a path for novel applications in real experimental data, enabling high-precision analyses across a wide range of imaging-based experiments.

Abstract

Accurate separation of signal from background is one of the main challenges for precision measurements across high-energy and nuclear physics. Conventional supervised learning methods are insufficient here because the required paired signal and background examples are impossible to acquire in real experiments. Here, we introduce an unsupervised unpaired image-to-image translation neural network that learns to separate the signal and background from the input experimental data using cycle-consistency principles. We demonstrate the efficacy of this approach using images composed of simulated calorimeter data from the sPHENIX experiment, where physics signals (jets) are immersed in the extremely dense and fluctuating heavy-ion collision environment. Our method outperforms conventional subtraction algorithms in fidelity and overcomes the limitations of supervised methods. Furthermore, we evaluated the model's robustness in an out-of-distribution test scenario designed to emulate modified jets as in real experimental data. The model, trained on a simpler dataset, maintained its high fidelity on a more realistic, highly modified jet signal. This work represents the first use of unsupervised unpaired generative models for full detector jet background subtraction and offers a path for novel applications in real experimental data, enabling high-precision analyses across a wide range of imaging-based experiments.

Figures (17)

• Figure 1: Schematic of jet production and detection in heavy-ion collisions. Red lines show heavy-ion beams. Yellow lines are charged particle tracks. Blue boxes are calorimeters measuring energy. Light-gray cones show the jet signals, which are difficult to identify due to the large underlying event activity (background) in these collisions.
• Figure 2: Schematic diagram of the UVCGAN-S framework (a) training and (b) signal extraction. UVCGAN-S translates a jet plus background calorimeter image (Domain $B$) into its two separated components (Domain $A$): the isolated jet signal and the underlying background. See \ref{['fig:stratified_arch']} for more detail about the UVCGAN-S's training and inference procedure.
• Figure 3: Jet $\eta$ position difference ($\Delta\eta = \eta^{\mathrm{sub}} - \eta^{\mathrm{real}}$) distribution (Left) for $R=0.5$ at $30<p_\mathrm{T}\xspace<40$ GeV. The right panel shows the root mean square (RMS) of $\Delta\eta$ (a measure of position resolution) as a function of the ground truth jet $p_\mathrm{T}$ for $R=0.5$, demonstrating superior performance of the UVCGAN-S.
• Figure 4: Jet transverse momentum response. Top panels show example distributions of $p_\mathrm{T}^\mathrm{sub}/p_\mathrm{T}^\mathrm{real}$ for $32<p_\mathrm{T}\xspace<34$ GeV bin for $R=0.2$ (Left) and $R=0.5$ (Right). Panels in the middle row are mean of $p_\mathrm{T}^\mathrm{sub}/p_\mathrm{T}^\mathrm{real}$ as a function of $p_\mathrm{T}^\mathrm{real}$ for $R=0.2$ (Left) and $R=0.5$ (Right). Bottom panels are resolution (standard deviation of $p_\mathrm{T}^\mathrm{sub}/p_\mathrm{T}^\mathrm{real}$) as a function of $p_\mathrm{T}^\mathrm{real}$ for $R=0.2$ (Left) and $R=0.5$ (Right).
• Figure 5: Jet reconstruction efficiency (Left) and fake rate (Right) as a function of the ground-truth jet $p_\mathrm{T}^\mathrm{real}$ for $R=0.5$. The statistical uncertainties are smaller than the marker sizes.