⚛️ phenomenology

Domain Adaptation against Background Sculpting in Anomaly Detection at the LHC

This paper proposes a domain-adaptation-based method to decorrelate anomaly scores from resonant mass, effectively mitigating background sculpting in weakly supervised anomaly detection at the LHC while preserving or even recovering sensitivity to new physics signals.

Original authors: Vincent Benne, Marie Hein, Michael Krämer, Humberto Reyes-Gonzalez, Philipp Soldin, Christopher Wiebusch

Published 2026-09-17

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Vincent Benne, Marie Hein, Michael Krämer, Humberto Reyes-Gonzalez, Philipp Soldin, Christopher Wiebusch

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, high-energy collisions of the Large Hadron Collider, physicists are constantly searching for the faintest whisper of new physics hidden within a roar of ordinary matter. They look for rare particles that might appear as a sudden, localized spike in the data, a resonance that stands out against a smooth, predictable sea of background events. To find these needles in the haystack, researchers have increasingly turned to machine learning, training computers to spot anomalies—strange patterns that don't fit the known rules of the universe. These algorithms assign a score to every collision event, flagging the ones that look most unusual. The goal is to set a threshold on this score, keeping the strange events and discarding the mundane ones, hoping to reveal a new particle hiding in the selected group.

However, there is a subtle trap in this approach. The computer's "strangeness" score can accidentally become linked to the mass of the particles involved in the collision. If the score is correlated with mass, selecting the strangest events doesn't just filter for new physics; it also reshapes the background data in a way that depends on mass. This creates artificial bumps or dips in the data that look exactly like the signals scientists are hunting for, or worse, they can hide real signals behind fake ones. This phenomenon, known as background sculpting, makes it incredibly difficult to tell if a bump in the graph is a discovery or just an artifact of the selection process.

A team of researchers from RWTH Aachen University and ETH Zurich has proposed a new method to untangle this knot, using a technique called domain adaptation. In their study, they treated the problem as a challenge of teaching a computer to ignore specific information it shouldn't care about. They designed a system where the machine learning model had to perform two tasks simultaneously: identifying the anomalous events and, at the same time, failing to distinguish between different regions of the background data. By forcing the model to lose the ability to tell apart data from the left and right sides of the mass spectrum, the researchers effectively stripped the anomaly score of its unwanted connection to the particle mass.

The team tested this approach using data from the LHC Olympics, a benchmark challenge for anomaly detection algorithms. They simulated millions of background collision events and injected a small number of signal events representing a hypothetical new particle. They applied their method to several different anomaly detection strategies, including one that uses a simple comparison between the signal region and the background, and more complex methods that estimate the background density. In every case, they compared the results of their new method against the standard approaches.

The results showed that the new technique successfully smoothed out the background distribution. When the researchers applied their domain adaptation, the artificial structures that usually appeared in the background data after selection largely disappeared. The background distribution remained smooth and predictable, which is essential for accurately estimating what the background should look like without any new physics. Crucially, this cleanup did not come at the cost of missing the signal. The method preserved the ability to detect the injected particles, maintaining high sensitivity even as it removed the confusing correlations.

In some specific scenarios where the input data had strong, natural correlations with the particle mass, the standard methods failed completely, producing distorted backgrounds and losing the ability to see the signal. In these difficult cases, the domain adaptation approach not only cleaned up the background but also recovered the lost sensitivity, allowing the detection of signals that would have otherwise been missed. The researchers found that a classification-based version of their method, which taught the model to confuse the left and right sides of the data, generally offered the best balance between cleaning the background and keeping the signal detection strong.

This work demonstrates that it is possible to use powerful machine learning tools for discovery without letting them distort the very data they are meant to analyze. By teaching the algorithms to be indifferent to certain variables, scientists can prevent the tools from creating their own false alarms. While the method requires careful tuning of how much emphasis to place on the "forgetting" task, the study suggests that this flexible approach could be combined with various existing search strategies. It offers a promising path forward for future searches at the collider, ensuring that when a bump appears in the data, it is a genuine sign of new physics rather than a trick of the selection process.

Technical Summary: Domain Adaptation against Background Sculpting in Anomaly Detection at the LHC

Problem Statement
In resonance searches for Beyond the Standard Model (BSM) physics at the Large Hadron Collider (LHC), weakly supervised anomaly detection (AD) has emerged as a powerful, model-agnostic tool. These methods typically construct an "anomaly score" to distinguish potential signal events from background. However, a critical challenge arises when the anomaly score correlates with the resonant mass ( $m_{JJ}$ ). This correlation induces "background sculpting": when events are selected based on the anomaly score, the background distribution in the resonant mass spectrum acquires artificial structures (bumps or dips) that mimic or obscure genuine resonant signals. This complicates standard background estimation techniques, which rely on fitting smooth sideband distributions to predict the background in the signal region. Previous solutions, such as quantile regression or limiting feature correlations, have been proposed, but a robust method to decorrelate the anomaly score from the resonant mass without sacrificing detection sensitivity remains a priority.

Methodology
The authors propose a novel solution using Domain Adaptation by Backpropagation (DAB) [14] to mitigate background sculpting. The core idea is to train the anomaly detection model such that the learned feature representation is discriminative for the main task (identifying anomalies) but invariant to the "domain," defined here by the resonant mass.

Architecture: The standard Multi-Layer Perceptron (MLP) classifier is split into a "feature extractor" ( $G_f$ ) and an "anomaly detector" ( $G_y$ ). A separate "domain network" ( $G_d$ ) is attached to the feature extractor via a gradient reversal layer.
Loss Function: The total loss is $L = L_y + \lambda \cdot L_d$ , where $L_y$ is the anomaly detection loss and $L_d$ is the domain loss. The gradient reversal layer ensures that while $G_d$ minimizes $L_d$ (learning to distinguish domains), the feature extractor $G_f$ is updated to maximize $L_d$ (preventing the features from retaining domain information).
Domain Tasks: Two specific domain tasks are tested using sideband data (regions outside the signal region):
1. Classification: Distinguishing between events from the left and right sidebands. Since both are pure background, the only distinguishing factor is their mass; preventing this classification forces decorrelation from mass.
2. Regression: Predicting the resonant mass directly from the features. Degrading this prediction reduces mass information in the representation.
Anomaly Detection Methods: The DAB approach is applied to four weakly supervised AD methods:
1. Idealized Anomaly Detector (IAD): Uses an idealized background template of pure signal-region background events.
2. CATHODE: Uses a conditional density estimator trained on sidebands to generate a background template.
3. CWoLa Hunting: Classifies signal-region events against sideband events without an explicit background model.
4. LaCATHODE: A previous method designed to mitigate sculpting by performing classification in a decorrelated latent space.

Experimental Setup
The study utilizes the LHC Olympics 2020 R&D dataset, consisting of QCD dijet background events and resonant $W'$ signal events ( $m_{W'} = 3.5$ TeV). Two feature sets are evaluated:

Baseline: Features with weak correlations to $m_{JJ}$ (jet masses, mass differences, subjettiness).
Baseline+ $\Delta R$ : Includes the angular distance between jets, introducing strong correlations with $m_{JJ}$ to stress-test the methods.

Performance is evaluated using:

Sculpting Severity: Measured by a binned $\chi^2$ /dof metric comparing the normalized mass distribution before and after selection. Values near unity indicate no sculpting.
Anomaly Detection Performance: Measured by the Significance Improvement Characteristic (SIC), $\epsilon_S / \sqrt{\epsilon_B}$ .

Key Results

Mitigation of Sculpting: Domain adaptation substantially reduces background sculpting across all tested methods. For the IAD and CATHODE methods, the classification domain task reduces the $\chi^2$ /dof to values close to unity for both feature sets, effectively removing artificial structures while preserving the anomaly detection performance (SIC).
Recovery of Sensitivity in CWoLa: For CWoLa Hunting with the strongly correlated Baseline+ $\Delta R$ features, the standard method fails completely (severe sculpting and near-zero sensitivity) because the classifier learns to distinguish signal regions from sidebands based on mass. Domain adaptation, particularly the regression task, successfully recovers anomaly detection sensitivity and significantly reduces sculpting, demonstrating the ability to restore performance when underlying assumptions are violated.
Comparison with LaCATHODE: While LaCATHODE effectively reduces sculpting, it often results in a slight reduction in anomaly detection sensitivity compared to the original CATHODE setup. The proposed DAB approach maintains sensitivity more closely to the original methods while achieving comparable decorrelation.
Task Comparison: The classification domain task generally offers the best compromise between sculpting reduction and signal sensitivity for IAD and CATHODE. The regression task is more effective at recovering sensitivity in the challenging CWoLa Baseline+ $\Delta R$ scenario but sometimes incurs a slight loss in sensitivity for other methods.

Significance and Claims
The paper claims that domain adaptation provides a flexible and complementary strategy to existing methods like LaCATHODE. Unlike LaCATHODE, which requires a specific latent-space construction, DAB acts directly on the learned feature representation and is not tied to a specific background template generation method. This flexibility suggests DAB could be combined with various background template approaches (e.g., TRANSIT, RADOT) and other neural network architectures.

The authors note practical limitations: the method requires tuning the hyperparameter $\lambda$ (loss weight), which currently relies on signal models (a challenge for truly model-agnostic analyses), and it is restricted to differentiable models (neural networks), excluding boosted decision trees. The study is limited to a single signal benchmark and uses a $\chi^2$ metric for sculpting rather than full sideband fits, but the results demonstrate that DAB can substantially reduce background sculpting while largely preserving anomaly detection performance, potentially broadening the range of features usable in resonance searches.

Technical Summary: Domain Adaptation against Background Sculpting in Anomaly Detection at the LHC

More like this