Calibration of electromagnetic shower features in the CMS calorimeter with machine-learning techniques

In the world of high-energy physics, scientists do not simply look at particles; they build them, at least in their minds. To understand what happens when protons smash together at nearly the speed of light, researchers rely on vast computer simulations. These digital models act as a mirror to reality, predicting how particles should behave, how they should scatter, and how they should leave their mark on a detector. If the simulation matches the real data collected by giant machines, scientists can trust the model to reveal the secrets of the universe. But if the simulation is slightly off—if it misrepresents the way a particle showers energy or interacts with the detector material—those small errors can grow into large mistakes, hiding new discoveries or creating false ones. The goal is always to make the digital mirror as perfect as possible, so that when a scientist sees a signal in the data, they know it is real and not just a glitch in the code.

This is the challenge faced by the CMS Collaboration, a team of physicists working with the Compact Muon Solenoid detector at CERN. In a recent study, they tackled a persistent problem: their computer simulations of electromagnetic showers—cascades of energy produced by electrons and photons—did not perfectly match the real-world data collected during proton-proton collisions in 2022. The discrepancies were subtle but significant, particularly in the shape of the energy deposits and how isolated the particles were from their surroundings. To fix this, the team turned to machine learning, not to replace their physics models, but to teach the simulations how to look more like the real thing. They developed two distinct methods to recalibrate the digital events, effectively training the computer to correct its own mistakes before the data was ever analyzed.

The researchers started with a clean, well-understood sample of events: the decay of a Z boson into an electron and a positron. By selecting these specific events, they could isolate a pure stream of electromagnetic showers, free from the noise of other particle interactions. They then compared the features of these showers in their simulation against the features observed in the actual data collected by the CMS detector. The data came from 26.7 inverse femtobarns of collisions at a center-of-mass energy of 13.6 tera-electronvolts. The simulation, while generally good, showed deviations in how the energy spread out across the detector crystals and how much energy leaked into surrounding areas. These differences meant that the computer's particle identification algorithms, which decide whether a signal is a genuine electron or a background mimic, were being fed slightly wrong information.

To bridge this gap, the team applied two different machine-learning techniques. The first approach, called reweighting, works by assigning a new importance score to every single simulated event. Imagine a classroom where the teacher asks students to raise their hands if they fit a certain description. If the simulation has too few students fitting that description compared to the real class, the teacher tells those few students to count as two or three people. In the same way, the machine learning classifier learned to distinguish between the simulation and the real data. It then calculated a ratio, effectively telling the simulation, "This event looks a bit like the real data, so count it more," or "This event looks too different, so count it less." This method adjusted the probability of each event occurring without changing the event itself, creating a continuous correction that could be applied across a complex, high-dimensional space of features.

The second method, known as normalizing flow, took a more direct route. Instead of just changing how much an event counts, this technique physically altered the values of the features within the simulation. It learned a mathematical transformation that could morph the distribution of the simulated showers into the distribution of the real showers. If a simulated shower was too narrow, the flow would stretch it; if it was too isolated, the flow would bring it closer to its neighbors. This process was like reshaping clay: the material remained the same, but its form was adjusted to match a mold. Crucially, this transformation was learned to be conditional on the kinematic properties of the event, such as the particle's momentum and position, ensuring that the correction was applied correctly regardless of where or how fast the particle was moving.

Both methods were tested on a separate set of data that the algorithms had not seen during their training. The results were striking. Before the correction, the simulated particle identification scores showed clear mismatches with the real data, particularly in the tails of the distributions where rare events occur. After applying either the reweighting or the normalizing flow corrections, the agreement between the simulation and the data improved dramatically. The shapes of the distributions aligned, and the discrepancies that had previously required large safety margins in scientific measurements were significantly reduced. The researchers found that both techniques successfully corrected the high-dimensional feature space, capturing not just individual variables but also the complex correlations between them.

However, the two methods had different strengths and limitations. The reweighting method was excellent at preserving the overall structure of the simulation while adjusting the weights, but it could sometimes struggle in regions where the simulation and data were very different, leading to large weight corrections that reduced the statistical power of the sample. In some cases, the team had to exclude certain features from the training to prevent these large fluctuations. The normalizing flow method, on the other hand, was able to handle large discrepancies more gracefully by directly reshaping the data, but it required that the corrected features be re-evaluated through the particle identification algorithm, adding a step to the process. Furthermore, while the reweighting method naturally preserved the laws of physics by leaving the event values untouched, the normalizing flow method had to be carefully monitored to ensure it did not create unphysical configurations, though the team found this risk to be negligible for their specific application.

The study demonstrated that these machine-learning techniques could be successfully applied to real collision data, a significant step forward from previous tests that relied only on simulated pseudo-data. By correcting the simulation to match the data, the researchers showed that they could reduce the systematic uncertainties that usually plague such measurements. This means that future analyses, such as those searching for the Higgs boson decaying into two photons, can be performed with greater precision. The work also highlighted the importance of identifying where simulations fail; the large corrections required by the algorithms pointed directly to specific areas in the detector modeling, such as the noise thresholds in the hadronic calorimeter, which the team was then able to fix for future iterations of the simulation.

Ultimately, this research provides a new toolkit for particle physicists. It shows that machine learning can do more than just classify particles; it can act as a calibration tool, teaching simulations to be more honest about the reality they are trying to describe. By offering continuous, unbinned corrections across high-dimensional spaces, these methods allow scientists to move beyond simple, binned adjustments that often miss the nuances of complex data. The success of these techniques in the CMS detector suggests that they will become a standard part of the analysis chain, helping to sharpen the focus of the search for new physics and ensuring that the signals scientists find are as clear and reliable as possible. The paper concludes that while the methods are powerful, they must be applied with care, understanding their specific limitations and the nature of the data they are correcting, to ensure that the digital mirror reflects the universe with the highest possible fidelity.