Enhancing searches for resonances with machine learning and moment decomposition

The Big Picture

Imagine you’re searching for a needle in a haystack, but the haystack itself is constantly shifting. Now imagine you train a powerful AI to spot needles, only to discover the AI has accidentally bent some straw into needle shapes. That’s roughly the problem facing physicists hunting for new particles at the world’s most powerful colliders.

The technique is called bump hunting: physicists look for unexpected spikes in a smooth spectrum of energy and mass. When a new particle exists, the lighter particles it decays into pile up at a characteristic energy, creating a visible spike above the otherwise smooth background. It’s how the Higgs boson was discovered, and the method dates back to the ρ meson in the 1960s. Today, bump hunts scan energies into the multi-TeV range, territory accessible only at massive accelerators like the LHC.

The trouble begins when physicists sharpen their searches using ML classifiers. A well-trained neural network may learn that certain collision features correlate with the total mass-energy of the event. Apply a score cutoff, and suddenly the background, which should be featureless, develops an artificial spike. A fake signal. A ghost.

Researchers from MIT and Lawrence Berkeley National Laboratory have developed a new framework called Moment Decomposition (MoDe) that sidesteps this problem, achieving higher sensitivity to new physics without corrupting the background estimate.

Key Insight: You don’t need a classifier to be completely independent of mass to avoid fake bumps. You just need it to avoid creating localized structures. MoDe exploits this looser requirement to build classifiers that are both more powerful and safe.

How It Works

The standard fix to bump-sculpting has been decorrelation: forcing the classifier to be statistically independent of mass, so it cannot use any mass information when deciding whether an event looks like a signal. Techniques like Distance Correlation (DisCo), adversarial networks, and planing all penalize mass correlations. They work, but they overpay.

By demanding complete independence, decorrelation handicaps the classifier. It cannot exploit any mass-related information, even information that would boost signal sensitivity without causing sculpting.

MoDe’s central insight is that this restriction is too strong. A classifier with a perfectly linear dependence on mass creates a sloped background after selection, but no localized bump. A bump hunt cares about bumps, not slopes. The researchers formalized this with a moment loss function, a training penalty that suppresses only the mass-dependent patterns that actually cause sculpting.

Here’s the machinery. The classifier output is expanded in terms of its Legendre polynomial moments, mathematical basis functions that decompose a distribution into constant, linear, quadratic, and higher-order components (like breaking a complex curve into simple building blocks). The loss function then selectively penalizes only those components that would sculpt the background:

• MoDe-0: penalizes any mass dependence, equivalent to standard decorrelation
• MoDe-1: allows linear mass dependence, penalizes higher moments
• MoDe-2: allows linear and quadratic dependence, even more flexible

Analysts can add further constraints: bound the slope of linear dependence, or restrict quadratic terms to be monotonic. These prevent dangerous background curvature even when the classifier exploits mass information.

The team validated MoDe on two test cases. The first was a synthetic model designed to cleanly illustrate each method’s behavior. The second was boosted hadronic W tagging, identifying highly energetic W bosons (force carriers of the weak nuclear interaction) decaying to quark pairs, distinguished from the overwhelming flood of jets produced by the strong force. This is the kind of analysis that matters for heavy-resonance searches at the LHC, where both ATLAS and CMS have deployed ML-based W taggers.

In both cases, MoDe-1 and MoDe-2 achieved better signal-to-background discrimination than fully decorrelated methods. Background distributions in the sideband regions (the portions of the mass spectrum away from a signal, used to estimate the background shape) remained smooth and usable throughout.

Why It Matters

For years, the implicit assumption in particle physics has been: if decorrelation prevents sculpting, then decorrelation is what we need. MoDe cracks that open and asks what you actually require. The answer is a weaker condition, and the gap between “sufficient” and “necessary” is where new sensitivity lives.

The payoff extends well beyond W tagging. Any bump hunt combining ML-enhanced selection with sideband background estimation could benefit: searches for dark matter mediators, heavy resonances, lepton flavor violation, and exotic nuclear signatures. The moment loss is fully differentiable and slots into standard training pipelines without modification. It scales to higher-order moments and multi-dimensional protected attributes. The authors also note that analogous ideas appear in domain adaptation and algorithmic fairness, and MoDe’s relaxed-constraint approach may inform those fields in turn.

Bottom Line: MoDe gives particle physicists a sharper tool for new-physics searches by replacing an overly conservative independence requirement with a precisely calibrated tolerance for controlled mass dependence. More flexibility, same safety guarantee.