DeepMET: Improving missing transverse momentum estimation with a deep neural network

DeepMET is a novel missing transverse momentum estimator developed by the CMS Collaboration that utilizes deep neural networks to assign weights to reconstructed particles, resulting in improved resolution, greater versatility across final states, and increased resilience to pileup compared to existing methods.

The Gist

The Invisible Ghost Hunter: How CERN is Using AI to Find "Missing" Particles

Imagine you are at a massive, high-speed bumper car arena. Hundreds of cars are zooming around, crashing into each other, and flying off in different directions. You are sitting in a control room, watching a giant screen that tracks every single car using sensors.

Suddenly, you notice something strange. You see a car get hit from the side with incredible force, sending another car spinning wildly to the left. But when you look at your sensors, there is no car there. No car hit it. No car is visible on the screen. Yet, the laws of physics say that if a car was hit, something must have caused it.

In the world of particle physics at CERN, these "ghost cars" are real. They are particles like neutrinos or mysterious dark matter that pass right through our detectors without leaving a trace. We can’t see them directly, so we have to play detective. We look at the "mess" they leave behind—the way they push visible particles around—and try to calculate exactly how much "invisible stuff" must have been there.

This invisible force is called Missing Transverse Momentum ($\vec{p}_T^{miss}$).

The Problem: The "Noisy Party" Effect

Calculating this invisible force is incredibly hard because the Large Hadron Collider (LHC) is a very "noisy" place.

Think of it like trying to calculate the force of a secret ghost hitting a billiard ball in the middle of a crowded, noisy nightclub. Not only are there the billiard balls you care about, but there are also hundreds of other people (called "pileup") bumping into each other, spilling drinks, and creating a chaotic mess of movement. If you try to add up all the movement in the room to find the "ghost," the noise from the crowd will give you a completely wrong answer.

The Solution: DEEPMET (The Super-Smart Bouncer)

For years, scientists used standard mathematical formulas to try and filter out this noise. It worked okay, but it wasn't perfect.

The CMS Collaboration has now introduced a new tool called DEEPMET. Instead of using a rigid formula, they built a Deep Neural Network—a type of Artificial Intelligence.

The Analogy:
If the old methods were like a basic calculator trying to subtract the noise, DEEPMET is like a world-class bouncer with superhuman intuition.

Instead of just looking at the total movement in the room, DEEPMET looks at every single person (every particle) individually. It asks:

• "Is this person moving like they belong to the main collision, or are they just a random person from the crowd (pileup) stumbling by?"
• "Is this particle a heavy, reliable 'car,' or is it just a tiny, jittery piece of debris?"

Based on these questions, DEEPMET assigns a "weight" to every particle. It gives a high weight to the particles that are clearly part of the main event and a very low weight to the "noise" from the crowd. It then adds up these weighted movements to reveal the true path of the "ghost."

Why This Matters

The paper shows that DEEPMET is a game-changer:

1. It’s Sharper: It improves the accuracy of our "ghost hunting" by 10% to 30%. It’s like upgrading from a blurry security camera to 4K high-definition.
2. It’s Tougher: It doesn't get confused by the "pileup" (the noisy crowd) nearly as much as previous methods.
3. It’s Versatile: It works whether we are looking for known particles like neutrinos or searching for the ultimate prize: Dark Matter.

By using AI to clean up the chaos, CERN scientists can now see the invisible more clearly than ever before, bringing us one step closer to understanding the hidden parts of our universe.

Technical Summary

Technical Summary: DEEPMET – Improving Missing Transverse Momentum Estimation with a Deep Neural Network

1. Problem Statement

In hadron collider experiments like those at the LHC, many critical physics processes (e.g., neutrino production in the Standard Model or dark matter candidates in Beyond-the-Standard-Model scenarios) involve particles that do not interact with the detector. These are inferred via missing transverse momentum ($\vec{p}_T^{\text{miss}}$), calculated as the negative vector sum of all reconstructed particle transverse momenta.

Measuring $\vec{p}_T^{\text{miss}}$ with high precision is extremely challenging due to:

• Detector limitations: Finite coverage, noise, and energy resolution.
• Reconstruction inefficiencies: Incomplete identification of particles.
• Pileup: Additional proton-proton interactions in the same or nearby bunch crossings that add "noise" momentum, degrading the resolution.
• Limitations of existing methods: Previous machine-learning approaches (like MVA-$\vec{p}_T^{\text{miss}}$) were often trained on specific physics processes, limiting their generalizability across different kinematic regions and event topologies.

2. Methodology

The CMS Collaboration introduces DEEPMET, a novel $\vec{p}_T^{\text{miss}}$ estimator based on a Deep Neural Network (DNN).

• Algorithm Architecture: Unlike global event-level estimators, DEEPMET operates on individual Particle-Flow (PF) candidates. The DNN processes 11 features per particle (8 continuous, such as impact parameters $d_{xy}, d_z$, PUPPI weights, and 5 kinematic variables; and 3 categorical, such as Particle ID, charge, and LPV association).
• Mechanism: The DNN assigns a unique weight ($w_i$) and two bias terms ($b_{i,x}, b_{i,y}$) to each reconstructed particle. The final $\vec{p}_T^{\text{miss}}$ is the negative vector sum of these weighted transverse momenta plus the bias contributions:
  \vec{p}_T^{\text{miss}}_x = -\sum_i (w_i p_{i,x} + b_{i,x}), \quad \vec{p}_T^{\text{miss}}_y = -\sum_i (w_i p_{i,y} + b_{i,y})
• Training: The model is trained using a Mean Squared Error (MSE) loss function on simulated $Z+\text{jets}$ and $t\bar{t}$ samples. This choice of training data ensures a wide range of jet multiplicities and $p_T$ distributions.
• Robustness (PV-Agnostic Version): To handle cases where the Leading Primary Vertex (LPV) is incorrectly identified (a common issue in high-pileup or high-$\eta$ environments), a "PV-Agnostic" version was developed. It uses a more robust vertex selection method to ensure stable performance in $W \to \mu\nu$ analyses.
• Calibration: To align simulation with real data, the authors employ a quantile-based calibration using $Z \to \mu\mu$ events to correct for detector asymmetries ($\phi$ dependence) and scale/resolution differences.

3. Key Contributions

• Particle-Level Weighting: Shifting the ML focus from global event variables to individual particle properties allows for better local information processing.
• Generalizability: Because the DNN does not rely on event-level topology, it generalizes well across diverse simulated processes (Higgs, Supersymmetry, etc.) and real collision data.
• Computational Efficiency: With only 4,541 trainable parameters, the model is lightweight, allowing for fast inference ($\sim 10$ ms per event) within the CMS software framework (CMSSW).
• Pileup Resilience: The inclusion of PUPPI weights and the specific DNN architecture make the estimator significantly more robust against pileup compared to traditional methods.

4. Results

• Resolution Improvement: DEEPMET improves the $\vec{p}_T^{\text{miss}}$ resolution by 10–30% compared to existing CMS estimators (PF $\vec{p}_T^{\text{miss}}$ and PUPPI $\vec{p}_T^{\text{miss}}$).
• Pileup Performance: In high-pileup regimes (e.g., $\sim 20$ reconstructed vertices), DEEPMET shows a massive advantage, outperforming PUPPI $\vec{p}_T^{\text{miss}}$ by up to 30% in resolution.
• Physics Process Versatility: Performance gains were validated across a wide spectrum of final states, including:
  • Standard Model processes ($Z+\text{jets}, W+\text{jets}, t\bar{t}$).
  • Higgs physics (Invisible Higgs, Di-Higgs).
  • Supersymmetry (Top squark and heavy neutralino production).
• Data-Simulation Agreement: After applying quantile-based calibrations, the DEEPMET performance in data shows excellent agreement with Monte Carlo simulations.

5. Significance

DEEPMET represents a significant advancement in the reconstruction of invisible physics signatures at the LHC. By providing a more precise and resilient measurement of $\vec{p}_T^{\text{miss}}$, it directly enhances the sensitivity of searches for Dark Matter and other Beyond-the-Standard-Model (BSM) physics. Furthermore, its efficiency and robustness make it a vital tool for the high-luminosity environment of future LHC runs, where pileup will become even more severe.