Source-linked AI summary
QCD or What?
Theo Heimel, Gregor Kasieczka, Tilman Plehn, Jennifer M Thompson
TL;DR
The paper addresses how to search for anomalous jet substructure without defining a signal or relying on labelled training data. It uses QCD-trained autoencoders on images or 4-vectors, with adversarial jet-mass de-correlation, and reports anomaly sensitivity alongside controllable background shaping and reduced phase-space-related systematics.
Problem
The paper asks whether jets can be searched for non-QCD patterns without predefined signal samples, amid limitations in training data, systematic control, and control over the learned physics question.
Method
The study trains autoencoders only on QCD jets using image or constituent 4-vector inputs, and uses an adversarial network to de-correlate jet mass.
Results
The autoencoder identifies anomalous heavy-resonance and exotic jets, while adversarial de-correlation produces controllable mass-sideband behavior with reported AUC values around 0.6 to 0.79 in tested examples.
Takeaways & Limitations
Training and applying the autoencoder on data in the same phase-space region supports an unsupervised search using jet-mass bump hunting and orthogonal follow-up analyses.
Takeaways & Limitations
The method can lose performance when jet-mass information is removed, and fake bumps from the network remain an uncertainty requiring tuning and control-region checks.
Abstract
from arXiv · showhide
Autoencoder networks, trained only on QCD jets, can be used to search for anomalies in jet-substructure. We show how, based either on images or on 4-vectors, they identify jets from decays of arbitrary heavy resonances. To control the backgrounds and the underlying systematics we can de-correlate the jet mass using an adversarial network. Such an adversarial autoencoder allows for a general and at the same time easily controllable search for new physics. Ideally, it can be trained and applied to data in the same phase space region, allowing us to efficiently search for new physics using un-supervised learning.
1 Introduction
The paper proposes unsupervised autoencoders trained only on QCD jets to identify non-QCD patterns, while adversarial de-correlation controls jet-mass information and related systematics. It develops image- and 4-vector-based approaches and aims to support same-phase-space training and analysis.
- 1 Introduction: Autoencoders trained on QCD jets search for non-QCD patterns without requiring a predefined signal sample.The networks compress jet information through a reduced latent representation to identify patterns not captured by that representation.
- 1 Introduction: The study considers both calorimeter jet images and constituent 4-vectors as input formats for autoencoder-based anomaly searches.The image approach uses a convolutional architecture, while the 4-vector approach uses a LoLa-like network.
- 1 Introduction: Deep learning jet analyses face limitations from training-data availability, systematic uncertainties, and control over the physics question answered.These constraints motivate unsupervised learning and methods that regulate which information the network uses.
- 1 Introduction: Adversarial de-correlation removes jet-mass information from the autoencoder, enabling control regions, sidebands, or smooth spectra for anomaly searches.The paper extends this idea to de-correlate any well-defined physics effect needed for a given analysis.
- 1 Introduction: Training and applying the autoencoder in the same phase-space region is intended to reduce uncertainties from relating simulation to data or background to signal regions.The proposed setup uses data for both training and analysis while preserving statistically independent samples.
2 Autoencoded QCD vs tops
Autoencoders trained on QCD jets identify anomalous top jets using either images or constituent 4-vectors, while adversarial training removes jet-mass dependence to support controlled searches. The LoLa autoencoder outperforms the image-based version before decorrelation, but image-based adversarial training provides the more stable mass-decorrelated approach.
- 2 Autoencoded QCD vs tops: An autoencoder trained only on QCD jets searches for non-QCD jets by compressing jet information through a reduced-dimensionality bottleneck.The study applies bottlenecks to both convolutional jet images and LoLa-like constituent 4-vectors.
- 2.2 LoLa: The LoLa autoencoder achieves a stable AUC around 0.92 with a bottleneck of at most 10 units for anomalous-top identification.The small bottleneck reflects information encoded by the CoLa/LoLa structure in physics-motivated features.
- 2.2 LoLa: The LoLa autoencoder performs better than the image-based autoencoder because its architecture extracts leading discriminating features efficiently with a smaller bottleneck.Both approaches are compared using ROC curves evaluated across independent test samples.
- 2.3 De-correlating the mass: Selecting the least QCD-like image-autoencoder jets sculpts a background mass peak near 200 GeV, motivating adversarial decorrelation of jet mass from the loss.The adversarial network is trained to produce a smooth QCD mass distribution independent of autoencoder loss.
- 2.3 De-correlating the mass: Increasing the adversarial strength improves background shaping but dilutes top enrichment and reduces anomaly-tagging performance as jet mass is removed from the discriminating information.The adversarial autoencoder retains a top-mass peak in the least QCD-like selection, enabling a controlled jet-mass shape analysis.
- 2.3 De-correlating the mass: Mass decorrelation is harder and less stable for the LoLa adversarial autoencoder, so the study focuses on image-based adversarial networks for subsequent analyses.The LoLa architecture is itself strongly focused on learning jet mass, which conflicts with decorrelating that observable.
- 2.3 De-correlating the mass: With 3% top contamination in training, a squeezed adversarial autoencoder retains top jets and their mass peak, achieving AUC 0.65 instead of 0.70 and recovering performance with λ = 3 · 10^-4.The remaining background shaping is similar to the pure-QCD adversarial case.
- 2.3 De-correlating the mass: Training and applying the network in the same phase-space region enables independent training and analysis samples, bump hunting in jet mass, and reduced leading systematics.A fake-bump propensity remains and must be controlled through hyperparameter tuning and additional data control regions.
3 Exotics in jets
The autoencoder is tested on heavy-scalar four-jet decays and dark showers, using adversarial mass de-correlation to search for signals whose substructure differs from QCD.
- 3 Exotics in jets: The signal-independent search targets heavy scalars decaying to four jets and dark showers, both chosen as exotic alternatives to benchmark top jets.The scalar replaces the top’s intermediate mass drop with higher subjet multiplicity, while dark showers add anomalous radiation and missing energy.
- 3.1 Scalar decay to jets: The image-based autoencoder reaches an AUC of 0.90 for scalar decays when trained on QCD jets without an adversary.This performance is comparable to the top benchmark, but the scalar case is harder because it lacks a second intermediate mass drop.
- 3.1 Scalar decay to jets: With 3% scalar signal contamination, the adversarial autoencoder enhances the signal near m_j = m_t in the 5% least QCD-like events, despite residual background shaping.The result demonstrates signal selection through the network output while retaining a distinctive mass feature.
- 3.2 Dark showers: Dark showers combine visible dijets, variable missing energy, and QCD radiation from heavy color-charged dark quarks, producing varied jet-mass signatures.The study examines both small and more mass-degenerate dark-meson masses.
- 3.2 Dark showers: Without an adversary, dark-shower models achieve AUC values of 0.78–0.79, while mass de-correlation lowers performance to around 0.6 but improves jet-mass control.The mass-degenerate model produces a broader, less pronounced peak among the 5% least QCD-like events.
4 Outlook
The outlook is an unsupervised anomaly-search framework trained and applied within the same phase-space region, with adversarial control of background shaping. It can identify multiple exotic jet signals, while reduced performance relative to specialized taggers is traded for broader applicability and lower systematics.
- 4 Outlook: Autoencoders trained on QCD or other background samples identify boosted top decays from images or 4-vectors, with reduced performance balanced by reduced systematic dependence.The approach is compatible with other jet representations and network architectures.
- 4 Outlook: Training and applying the network in the same phase-space region yields similar jet-mass distributions across loss-function slices, enabling controlled bump hunts.Top decays populate the least QCD-like slices and form a distinct jet-mass peak.
- 4 Outlook: With signal contamination, a more restrictive adversarial autoencoder still classifies top jets as least QCD-like and preserves their distinctive mass peak.The selected non-QCD slices remain usable for searching the peak.
- 4 Outlook: The adversarial autoencoder also extracts heavy-scalar four-quark decays and dark showers, allowing one network to search for multiple signals in the same phase-space region.The universal network structure is proposed as a basis for externally testing whether specific models would be flagged as anomalies.