Precision calibration of calorimeter signals in the ATLAS experiment using an uncertainty-aware neural network

This paper presents a Bayesian neural network approach for the multi-dimensional calibration of ATLAS calorimeter topo-clusters, which outperforms standard methods while providing robust uncertainty estimates that contribute to reducing systematic errors.

The Gist

Imagine the Large Hadron Collider (LHC) as the world's most powerful particle accelerator, smashing protons together to create a shower of new particles. The ATLAS experiment is a massive, high-tech camera designed to take pictures of these collisions. However, instead of a single lens, ATLAS uses a "calorimeter"—a giant, layered sandwich of detectors that acts like a cosmic rain gauge. When particles hit this sandwich, they leave behind energy deposits, which the machine reads as electrical signals.

The problem? The "rain gauge" isn't perfect. It's like a scale that weighs a feather differently than a brick, even if they have the same mass. In physics terms, the detector responds differently to different types of particles (like electrons vs. protons). To get the true energy of a particle, scientists have to apply a "calibration"—a mathematical correction factor.

For years, ATLAS used a standard, rule-based method (called LCW) to apply these corrections. It worked, but it was a bit clunky, like using a ruler with only inch marks to measure something that needs millimeter precision. It also couldn't easily tell you how sure it was about its measurement.

This paper introduces a new, smarter way to calibrate these signals using Artificial Intelligence (AI), specifically a type of "Bayesian Neural Network" (BNN). Here is how the paper explains it, using simple analogies:

1. The Old Way vs. The New Way

• The Old Way (LCW): Imagine you are trying to guess the weight of a mystery box. The old method uses a lookup table. If the box is red and small, you look up "Red/Small" in a book and find a correction factor. If the box is red and medium, you look up "Red/Medium." This creates "steps" in your data. If a box is right on the edge between "Small" and "Medium," the correction might jump suddenly, which isn't physically realistic.
• The New Way (BNN): The new AI method doesn't use a lookup table. Instead, it learns a smooth, continuous curve. It understands that a "medium-small" box should have a correction factor somewhere between the two, not a sudden jump. It looks at many features of the box (size, color, texture, where it was found) all at once to make a single, smooth prediction.

2. The "Confidence Meter" (Uncertainty)

This is the paper's biggest innovation. Standard AI models give you an answer (e.g., "The energy is 50 GeV"), but they don't tell you if they are guessing or if they are 100% sure.

The Bayesian Neural Network is like a weather forecaster who doesn't just say "It will rain," but also says, "It will rain, and I'm 90% sure, but there's a 10% chance I'm wrong because the sensors are acting up."

• Statistical Uncertainty: This is the "I need more data" feeling. If the AI has only seen 10 examples of a specific type of particle, it's less sure. If it sees a million, it becomes very sure.
• Systematic Uncertainty: This is the "I can't be more sure even with more data" feeling. This happens if the detector itself is noisy, or if the physics is inherently chaotic (like a pile of sand shifting). The AI learns to recognize these "messy" situations and raises a red flag, saying, "My answer might be off because the signal here is confusing."

3. How They Tested It

The scientists didn't just trust the AI; they put it through a rigorous "driving test."

• The Simulator: They used super-computers to simulate millions of particle collisions (Monte Carlo simulations). They knew the "true" energy of every particle because they created the simulation.
• The Comparison: They compared the new AI calibration against:
  1. The old standard method (LCW).
  2. A different type of AI (a standard Deep Neural Network).
  3. A "Repulsive Ensemble" (a second, completely different AI method designed to double-check the first one's confidence levels).

4. The Results

• Better Accuracy: The new BNN method was more accurate than the old standard, especially for low-energy particles. It smoothed out the "steps" and reduced errors.
• The Confidence Check: The "confidence meter" worked. When the AI was unsure (for example, when the detector signal was messy due to "pile-up"—when too many collisions happen at once), the uncertainty numbers went up.
• Agreement: The two different AI methods (BNN and the Repulsive Ensemble) agreed with each other. They both flagged the same "tricky" spots where the data was noisy. This proved the uncertainty numbers weren't just random glitches in the code; they were real reflections of the data's difficulty.

5. Why It Matters (According to the Paper)

The paper claims this method allows physicists to:

• Get a more precise measurement of energy.
• Know exactly when a measurement is shaky.
• Use this "confidence score" to filter out bad data before building complex physics models (like reconstructing jets or measuring missing energy).

In summary: The paper presents a new, AI-driven "smart ruler" for the ATLAS detector. It not only measures particle energy more smoothly and accurately than the old ruler but also comes with a built-in "confidence meter" that tells scientists exactly how much they should trust each individual measurement. This helps them separate the clear signals from the noisy background noise of the universe.

Technical Summary

Technical Summary: Precision Calibration of Calorimeter Signals in the ATLAS Experiment using an Uncertainty-Aware Neural Network

Problem Statement
The ATLAS experiment at the Large Hadron Collider (LHC) utilizes non-compensating calorimeters where the signal response to hadronic showers is systematically lower than that to electromagnetic showers. The standard approach to correct this involves a "Local Hadronic Calibration" (LCW) sequence, which applies a four-step procedure including cluster classification, local cell-signal weighting, and corrections for out-of-cluster effects. However, this standard method relies on multi-dimensional lookup tables derived from single-particle simulations. This approach averages correlations between observables and introduces discontinuities at bin edges, potentially limiting the precision of the calibration. Furthermore, the standard LCW does not provide per-cluster uncertainty estimates, which are crucial for propagating systematic uncertainties in complex event reconstruction (e.g., missing transverse momentum or jet substructure).

Methodology
This paper presents a novel calibration framework utilizing a Bayesian Neural Network (BNN) to learn a continuous, multi-dimensional calibration function for topo-clusters (clusters of topologically connected calorimeter cells).

• Training Target: Instead of predicting the deposited energy ($E_{dep}^{clus}$) directly, the network learns the response ratio $R_{clus}^{EM} = E_{clus}^{EM} / E_{dep}^{clus}$, where $E_{clus}^{EM}$ is the signal at the electromagnetic scale and $E_{dep}^{clus}$ is the truth-level energy deposited by primary particles within the cluster.
• Input Features: The network utilizes a feature set ($X_{clus}$) of 15 observables, including kinematic variables ($E_{clus}^{EM}$, rapidity), signal characteristics (significance, timing, compactness), shower shape parameters (longitudinal and lateral dispersion), and environmental variables (pile-up measures $N_{PV}$ and $\mu$).
• Network Architecture:
  • Bayesian Neural Network (BNN): The weights of the network are treated as probability distributions rather than fixed values. During inference, the network parameters are sampled $N=50$ times to generate an ensemble of predictions. This allows the model to output not only a central calibration value but also statistical and systematic uncertainty components. The loss function employs a Gaussian mixture model (3 modes) to accommodate deviations from a simple Gaussian likelihood and includes a Kullback-Leibler divergence term for regularization.
  • Repulsive Ensemble (RE): To validate the BNN uncertainty predictions, an alternative model using repulsive ensembles was trained. This method forces an ensemble of identically configured networks to explore the loss landscape differently by introducing a repulsive force between network members during training.
• Dataset: The models were trained and tested on full Monte Carlo simulations of proton-proton collisions at $\sqrt{s} = 13$ TeV (Run 2), specifically using multijet final states. The dataset includes approximately 14.5 million topo-clusters, with 60% used for training.

Key Results

• Calibration Performance: The BNN-derived calibration demonstrates superior performance compared to the standard LCW and a previously studied Deep Neural Network (DNN) approach.
  • Linearity: The BNN achieves a median deviation from linearity of $|\langle \Delta E \rangle_{med}| \lesssim 2\%$ for deposited energies $E_{dep}^{clus} \gtrsim 500$ MeV. This represents a significant improvement in the low-energy regime compared to the DNN, which requires $E_{dep}^{clus} \gtrsim 1$ GeV to reach similar accuracy.
  • Resolution: The relative local energy resolution is significantly improved over both the standard LCW and EM scales across all evaluated pile-up conditions ($N_{PV}$ and $\mu$). The BNN shows a flatter dependence on pile-up activity compared to the DNN, particularly at higher pile-up levels.
• Uncertainty Predictions:
  • The BNN successfully learns to predict uncertainties for each topo-cluster. These are decomposed into statistical (reducible with more data) and systematic (intrinsic to the model/data limitations) components.
  • A comparison between the BNN and the RE models shows a high degree of correlation in their uncertainty predictions (agreement within ~10%), suggesting that the predicted uncertainties reflect genuine data limitations and detector signal characteristics rather than artifacts of a specific network architecture.
  • Large predicted uncertainties are observed for topo-clusters in complex transition regions (e.g., Tile gap scintillators) and those dominated by pile-up, correctly identifying regions where the signal-to-truth relationship is less deterministic.

Significance and Claims
The paper claims that the application of a Bayesian Neural Network offers a robust method for calibrating ATLAS calorimeter signals that surpasses standard lookup-table methods in both precision and continuity. By learning a smooth, multi-dimensional function, the BNN preserves correlations between observables that are lost in binned approaches.

Crucially, the work demonstrates that uncertainty-aware machine learning can provide per-cluster uncertainty estimates that are interpretable as contributions to the systematic uncertainty of the calibration. The agreement between the BNN and the independent Repulsive Ensemble method validates the reliability of these uncertainty predictions. The authors posit that these uncertainties can be utilized to improve signal quality through topo-cluster selection and to propagate uncertainties in "bottom-up" calibration schemes for complex final states, such as soft hadronic recoil and jet substructure, thereby enhancing the overall precision of LHC measurements. The study remains focused on the validation of the method within Monte Carlo simulations, noting that the application to experimental data requires further performance studies to ensure feature correlations are well-modeled.