Electron Identification using Machine Learning in the MPD Experiment at NICA

In the heart of matter, where atoms dissolve into a soup of their smallest building blocks, lies a state of existence that once filled the entire universe just moments after the Big Bang. Scientists call this the Quark-Gluon Plasma, a seething, ultra-hot fluid where protons and neutrons melt apart into their constituent quarks and gluons. To study this primordial state, researchers recreate it in the lab by smashing heavy atomic nuclei together at incredible speeds. These collisions generate extreme temperatures and pressures, briefly forming a tiny droplet of this early-universe matter. However, to understand what happened inside that droplet, scientists must be able to see the specific particles that escape the explosion. Among the most important messengers are electrons, which carry a clean signal of the conditions inside the fireball. But finding these electrons is like trying to spot a single silver needle in a haystack of gold ones, because the collisions produce billions of other particles that look very similar.

At the Nuclotron-based Ion Collider fAcility in Russia, a massive new experiment called the Multi-Purpose Detector is being built to catch these collisions. The facility is designed to smash ions together at energies that allow scientists to explore how matter behaves when it is packed with a high density of protons and neutrons, a condition similar to the core of a neutron star. The detector is a large, barrel-shaped machine filled with layers of sensors that track the paths and energies of particles as they fly out from the collision point. One of the most critical tasks for the scientists working with this machine is to distinguish the rare electrons from the overwhelming background of other charged particles, such as pions and protons. If they cannot separate these signals cleanly, the physics they hope to uncover remains hidden.

The researchers behind this study focused on a specific challenge: how to identify electrons with the highest possible accuracy without losing too many of them in the process. Traditionally, scientists have used a method based on simple, one-dimensional rules. They would look at a particle's speed or its energy loss and draw a line, keeping everything on one side and discarding the rest. While this approach is straightforward, it is blunt. To ensure they are not keeping the wrong particles, they must set the rules very strictly, which inevitably throws away many of the correct electrons along with the bad ones. The team wanted to know if they could do better by using machine learning, a type of computer program that can learn to recognize complex patterns in data rather than following rigid, pre-set rules.

To test this idea, the team created a detailed computer simulation of the detector's response to collisions between heavy bismuth nuclei. They generated a massive dataset of simulated particle tracks, labeling each one as either an electron or a background particle. They then taught two different types of machine learning algorithms to sort through this data. The first was a neural network, a system modeled after the human brain that learns by adjusting connections between layers of simple processing units. The second was a boosted decision tree, which works by asking a long sequence of yes-or-no questions to narrow down the identity of a particle. Both systems were fed a variety of measurements, including the particle's momentum, how much energy it lost as it passed through the detector gas, and how its speed compared to its energy.

The researchers found that the machine learning approaches significantly outperformed the traditional method. When they used the standard one-dimensional rules, the system could achieve high purity, meaning the electrons it kept were almost certainly real, but it only managed to find about half of the total electrons available. The machine learning models, however, were able to find roughly 50 percent more electrons while maintaining that same high level of purity. This improvement was not uniform across all speeds; the algorithms had to be taught differently depending on how fast the particles were moving. By training separate models for different ranges of momentum, the team ensured that the system remained accurate even for the fastest particles, where the traditional methods tend to fail.

One of the most interesting discoveries was how the two machine learning models approached the problem differently. The neural network tended to rely heavily on the ratio of a particle's energy to its momentum, a measurement that is particularly useful for identifying electrons at high speeds. The decision tree, on the other hand, frequently used the measurement of energy loss in the gas, regardless of the particle's speed. This difference suggests that the two algorithms found different ways to distinguish the signal from the noise, yet both succeeded where the old method struggled. The result is a tool that allows physicists to see more of the rare electrons escaping the collision, effectively doubling the amount of usable data they can gather for their analysis.

This work represents a crucial step forward for the Multi-Purpose Detector as it prepares to begin operations. The machine learning techniques described here are not just theoretical exercises; they are practical tools that will be used to analyze real data once the detector is fully commissioned. By improving the ability to identify electrons, the experiment will be able to study the properties of the Quark-Gluon Plasma with much greater precision. The researchers note that while these results come from simulations, they provide a strong foundation for the actual data analysis. The ability to extract more signal from the same amount of data means that the experiment will be able to answer fundamental questions about the nature of matter and the early universe with a clarity that was previously out of reach.