⚛️ high-energy experiments

A machine-learning method for the estimation of the background from jets misidentified as hadronic tau lepton decays

The paper presents MUFFIN, a machine-learning method using boosted decision tree reweighting to estimate backgrounds from jets misidentified as hadronic tau leptons in CMS data, which significantly improves kinematic modeling and reduces associated systematic uncertainties by up to 70% compared to traditional low-dimensional parametrisation techniques.

Original authors: CMS Collaboration

Published 2026-10-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: CMS Collaboration

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 that occur inside the Large Hadron Collider at CERN, scientists are constantly searching for the subtle fingerprints of the universe's most fundamental particles. Among these, the tau lepton holds a special place. It is the heaviest member of the electron family, a particle so massive that it decays almost instantly after being created. When it decays into other particles, it often leaves behind a spray of hadrons, a specific type of particle that includes protons and neutrons. Detecting these sprays is crucial because they serve as a key signature for understanding the Higgs boson, the particle that gives other matter its mass, and for hunting down entirely new physics that lies beyond our current theories. However, finding these tau sprays is like trying to spot a specific type of bird in a dense forest; the forest is filled with ordinary jets of particles created by the collision of protons, and these jets can look remarkably similar to the tau sprays the scientists are seeking. Distinguishing the genuine signal from this overwhelming background of impostors is one of the most difficult challenges in modern particle physics.

For years, researchers have relied on a method known as the "Fake Factor" technique to estimate how often these ordinary jets trick their detectors into thinking they are tau leptons. This approach works by looking at specific regions of data where jets are known to be present and calculating a simple ratio to predict how many will appear in the regions where scientists are searching for new physics. While this method has been effective, it operates on a set of simplifying assumptions. It treats the complex behavior of particles as if it were a flat, two-dimensional map, ignoring the intricate, multi-layered relationships between different properties of the collision. As experiments have become more precise and the data more abundant, these simplifications have begun to show their limits, leading to uncertainties that can obscure the very signals scientists are trying to find.

A new study by the CMS Collaboration at CERN introduces a more sophisticated solution to this problem, utilizing machine learning to create a more accurate picture of the background noise. The team developed a technique they call MUFFIN, which stands for MUltivariate Fake-Factor INference. Instead of relying on simple, low-dimensional ratios, this new method uses a type of artificial intelligence known as a boosted decision tree to learn directly from the collision data itself. The system is trained to recognize the subtle, multidimensional patterns that distinguish a genuine tau lepton from a jet that merely mimics one. By analyzing a vast array of characteristics simultaneously—such as the energy of the particles, their direction, and the overall shape of the event—the algorithm constructs a detailed, three-dimensional map of how often jets are misidentified. This allows the researchers to apply a unique correction factor to every single particle candidate, rather than applying a broad average to a whole group.

The researchers tested this new approach using a massive dataset of proton-proton collisions recorded by the CMS detector in 2022 and 2023, corresponding to an integrated luminosity of 62.4 inverse femtobarns at a collision energy of 13.6 tera-electronvolts. They compared the performance of their machine-learning method against the traditional Fake Factor technique across various scenarios, including searches for the Higgs boson and studies of heavy new particles. The results showed a clear and significant improvement. The new method was able to model the behavior of the background jets with much greater fidelity, particularly in regions where the particles have very high energy, which are often the most critical areas for discovering new physics. In the traditional method, the simplifying assumptions often led to discrepancies between the predicted background and what was actually observed, requiring researchers to add large safety margins to their calculations to account for the uncertainty.

With the MUFFIN method, these discrepancies were dramatically reduced. The study found that the systematic uncertainty associated with the misidentification of jets as tau leptons was reduced by up to 70 percent in certain channels. This reduction is not merely a statistical detail; it fundamentally changes the sensitivity of the experiments. By tightening the constraints on the background noise, the researchers can now see more clearly into the data, making it easier to spot the faint signals of new particles or rare processes that might otherwise be lost in the noise. The method proved robust across different types of collision events and was shown to work effectively whether the tau leptons were produced in pairs or alongside other particles like muons.

The success of this approach demonstrates that machine learning can be successfully integrated into the core of experimental physics without losing the rigorous, data-driven nature of the field. The MUFFIN method does not rely on theoretical simulations to guess how the background behaves; instead, it learns the answer directly from the real-world data collected by the detector. This makes the results more reliable and the workflow more streamlined, removing the need for complex, multi-step corrections that were previously required to fix the shortcomings of older methods. As the Large Hadron Collider continues to collect data, this technique offers a powerful tool for future discoveries, allowing physicists to push the boundaries of their knowledge with greater confidence and precision. The ability to distinguish the signal from the noise with such clarity is a vital step forward in the ongoing quest to understand the fundamental laws that govern our universe.

Technical Summary: MUFFIN – A Machine-Learning Method for Jet-to-Tau Background Estimation

Problem Statement
In analyses involving hadronically decaying tau leptons ( $\tau_h$ ) at the CMS experiment, a primary experimental challenge is the background arising from quark- and gluon-initiated jets that are misidentified as $\tau_h$ candidates ( $jet \to \tau_h$ ). These backgrounds, originating largely from QCD multijet production, $W$ +jets, and top quark pair ( $t\bar{t}$ ) processes, mimic the complex experimental signature of $\tau_h$ decays. The standard approach for estimating this background is the "Fake Factor" (FF) method, which relies on low-dimensional parametrisations of transfer factors (TFs) derived from data-driven control regions. However, the FF method suffers from inherent limitations: it assumes the misidentification probability factorises across a limited set of kinematic variables, neglecting correlations between object-level and event-level quantities. This leads to statistical limitations in sparsely populated bins, systematic uncertainties from missing variable dependencies, and the need for complex, multi-step correction workflows that can reduce the overall precision and sensitivity of physics analyses, particularly in high- $p_T$ regions and multivariate discriminants.

Methodology: The MUFFIN Approach
The paper introduces MUltivariate Fake-Factor INference (MUFFIN), a machine-learning-based alternative to the FF method. MUFFIN generalises the background estimation paradigm by replacing explicit low-dimensional parametrisations with multidimensional, per-object reweighting functions learned directly from observed data.

Core Mechanism: The method employs Boosted Decision Trees (BDTs), implemented via the XGBOOST framework, trained to distinguish between $\tau_h$ candidates that pass nominal identification requirements ("pass") and those that fail nominal requirements but pass a looser working point ("fail").
Training Strategy: BDTs are trained in "Determination Regions" (DRs) enriched in specific background processes (QCD multijet, $W$ +jets, and $t\bar{t}$ ). The training inputs include local object properties (e.g., $\tau_h$ decay mode, $p_T$ , $\eta$ , $\phi$ , and the ratio of jet $p_T$ to $\tau_h$ $p_T$ ) and global event variables (e.g., total jet multiplicity, number of $b$ -tagged jets). Crucially, the method is fully data-driven for the targeted jet $\to \tau_h$ component; no simulation of the specific jet $\to \tau_h$ process is used in the training or subtraction steps.
Weight Derivation: The BDT outputs are calibrated using temperature scaling to ensure they represent well-calibrated class probabilities. A continuous, multidimensional transfer factor, $w_{\text{MUFFIN}}(z)$ , is constructed as the ratio of the probability of passing to the probability of failing, $P_{\text{pass}}(z) / P_{\text{fail}}(z)$ , where $z$ represents the full feature vector. Genuine $\tau$ contributions from simulation are subtracted to isolate the jet misidentification rate.
Validation: The method's performance is assessed in "Validation Regions" (VRs) orthogonal to the DRs and Signal Regions (SRs). The analysis utilises proton-proton collision data collected by CMS at $\sqrt{s} = 13.6$ TeV with an integrated luminosity of 62.4 fb $^{-1}$ (2022–2023).

Key Contributions

Multidimensional Reweighting: MUFFIN captures correlations between multiple kinematic and topological variables that the traditional FF method cannot efficiently parametrise due to its reliance on fixed binning and factorisation assumptions.
Data-Driven Robustness: The method maintains a fully data-driven nature for the dominant jet $\to \tau_h$ component, avoiding reliance on simulation for the specific misidentification probability of jets.
Unified Workflow: It provides a more reproducible and streamlined background estimation workflow, reducing the complexity associated with defining multiple parametrisations and correction steps.
Algorithm Independence: While demonstrated using the DEEPTAU v2.5 identification algorithm, the MUFFIN framework is designed to be adaptable to future identification algorithm developments.

Results
Performance studies comparing MUFFIN against the standard FF method demonstrate significant improvements in background modelling:

Improved Modelling: MUFFIN achieves superior modelling of key kinematic observables, including visible mass ( $m_{\text{vis}}$ ), total transverse mass ( $m_{\text{total}}_T$ ), and the angular separation ( $\Delta R$ ) between decay products. This improvement is particularly pronounced in the $\eta$ and decay mode distributions, where the FF method's marginalisation of these variables leads to residual mismodelling.
Systematic Uncertainty Reduction: The improved modelling translates directly into a reduction of systematic uncertainties associated with the jet $\to \tau_h$ background. In the $\tau_h\tau_h$ channel, the systematic uncertainty is reduced by up to 70% in high- $p_T$ tails and multivariate discriminant regions. In the $\tau_\mu\tau_h$ channel, reductions of approximately 10% are observed, with larger gains expected as data samples increase.
Closure Tests: In validation regions, MUFFIN consistently yields lower mean absolute Poisson pulls ( $\langle|z|\rangle$ ) compared to the FF method across various observables and analysis categories, including STXS (Simplified Template Cross Section) categories and high-mass resonance search regions.
Specific Channels: The method was applied to the fully hadronic $\tau_h\tau_h$ channel (where jet $\to \tau_h$ constitutes ~~75% of the background) and the semileptonic $\tau_\mu\tau_h$ channel (~~35% background), showing robust performance in both.

Significance and Claims
The paper claims that MUFFIN provides a robust and scalable method for background estimation in precision measurements and searches for physics beyond the Standard Model (BSM) involving $\tau_h$ final states. By learning the multidimensional dependence of the misidentification probability directly from data, MUFFIN addresses the limitations of low-dimensional parametrisations, particularly in kinematic regions of high sensitivity (e.g., high- $p_T$ tails and multivariate discriminant outputs) where the FF method often struggles. The author states that this approach enables a reduction in systematic uncertainties without introducing instabilities in well-modelled regions, thereby enhancing the sensitivity of future CMS analyses. The method is presented as a natural framework for improving background modelling precision as analyses increasingly exploit fine-grained categorisation and multidimensional phase space information.

More like this