Back

HEPHY wins FAIR Universe Higgs Uncertainty Challenge

A team from HEPHY, bringing together experts from the CMS and machine learning groups, has been named co-winner of the FAIR Universe Higgs Uncertainty Challenge – one of the first large-scale benchmarks for machine-learning-based inference techniques in collider physics. Among more than twenty international teams, our "GOLLUM" algorithm delivered top-tier performance on a detailed simulation of the Higgs boson decaying to a pair of tau leptons, earning the full $2,000 prize for its innovative approach.

 06.05.2025
Visible mass of the tau-lepton system. The top panel shows the distribution in simulated data. The middle panel shows that uncertainties are too large to measure the small signal process. The important information is in the bottom pad where our method reduces uncertainties drastically and a clear peak emerges.

What was the challenge?

The FAIR Universe Higgs Uncertainty Challenge asked participants to measure the rate of the pp → H → ττ process (cross-section) in the single-lepton channel, based on high-dimensional data comprising 28 kinematic and event-shape features. Teams had to contend with both statistical and systematic uncertainties, including jet and τ-lepton energy calibration as well as uncertainties in the rates of background processes. Most competitors, including the winning teams, turned to novel machine-learning methods rather than traditional histogram-based analyses.

How did we win?

Our GOLLUM algorithm combines three key ingredients to unlock the full potential of unbinned inference. First, fast surrogate models capture the known analytic dependencies of the process, such as the scaling with the signal strength and the background rates, without the need for approximation. Second, a calibrated multiclassifier uses high-dimensional event features to estimate per-event probabilities for signal- and background categories. The key here is an iterative isotonic regression step ensuring accurate predictions even across large class imbalances. Third, a parametric neural network learns how these class probabilities shift under systematic variations, yielding a smooth model across all model parameters. Together, these components form a fast and robust inference pipeline that significantly sharpens the measurement of the H → ττ signal strength.

The payoff

When applied to the unbinned dataset, our method achieves a substantial reduction in systematic uncertainty, clearly visible in the likelihood curves and reconstructed mass distributions. Even though the H → ττ signal is a small component of the dataset, the improved modeling leads to a highly significant bump in the lowest mass bins and a much narrower likelihood curve for the signal strength μ – decisively outperforming the traditional binned reference analysis. That improvement secured GOLLUM a tie for first place.

What’s next?

This result underscores the power of modern machine learning in pushing the precision frontier at the LHC. Several new HEPHY projects are currently starting in this direction. More importantly, it demonstrates the impact of close collaboration between domain physicists and machine-learning experts – a human model for future breakthroughs in precision measurements and searches for new physics.