⚛️ high-energy experiments

Validating a BDT-based Electron-Positron Identification Algorithm at CLAS12 with Experimental Data

This paper presents and validates a machine-learning-based Boosted Decision Tree algorithm for the CLAS12 experiment that effectively identifies electrons and positrons while significantly reducing charged-pion contamination, achieving over 90% lepton retention in simulated samples and demonstrating reliability on experimental data.

Original authors: Mariana Tenorio Pita, Pierre Chatagnon, Richard Tyson

Published 2026-09-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mariana Tenorio Pita, Pierre Chatagnon, Richard Tyson

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 subatomic world, scientists often act as cosmic detectives, trying to reconstruct the story of a collision by examining the debris that flies out. To do this, they need to know exactly what each piece of debris is. Is it an electron, a fundamental particle of light and electricity? Or is it a pion, a heavier, more common particle that looks very similar to an electron when it zips through a detector at high speed? This distinction is critical because if a scientist mistakes a pion for an electron, the entire calculation of the physical event falls apart. The challenge is that these particles move incredibly fast and leave behind only faint, overlapping traces of energy. To separate the true signal from the background noise, researchers rely on massive machines called particle accelerators, which smash particles together, and enormous detectors that record the aftermath. The goal is always the same: to see the universe clearly, without the blur of mistaken identity.

At the Thomas Jefferson National Accelerator Facility, a team of researchers faced this exact problem with the CLAS12 detector, a massive instrument designed to study the structure of matter. They needed a way to tell electrons and their antimatter counterparts, positrons, apart from a sea of charged pions. The detector was already quite good at this task, but it struggled when particles moved at certain speeds, specifically when they were fast enough to pass through a specific type of light detector but slow enough to still look like pions in the energy sensors. In these tricky zones, the detector would occasionally mistake a pion for a positron, creating a false signal that could ruin sensitive measurements. The researchers set out to build a smarter filter, one that could learn the subtle differences between these look-alike particles and clean up the data.

To solve this, the team turned to a branch of artificial intelligence known as machine learning, specifically using a method called a boosted decision tree. Imagine a series of questions asked in a row, where each answer helps narrow down the possibilities. The computer was trained on millions of simulated events, learning to recognize the unique "fingerprint" left by an electron or positron compared to a pion. Unlike a simple rule that says "if energy is high, it is an electron," this system looked at a complex combination of factors. It examined how much energy the particle deposited in different layers of a calorimeter, which is a device that stops particles and measures their energy, and it analyzed the shape of the energy shower the particle created. Electrons tend to spread their energy out in a tight, compact pattern, while pions scatter it more loosely. The computer learned to spot these patterns with incredible precision.

The team developed two versions of this smart filter. One version used a smaller set of clues, while the other used a more comprehensive list that included the particle's speed and direction. They tested both models rigorously. First, they ran them against the simulated data they used for training to ensure the logic held up. Then, they applied the filters to real data collected from the accelerator. The results were striking. The new algorithm successfully kept more than 90 percent of the true electrons and positrons in the sample, which is crucial because you do not want to throw away the real data you are looking for. At the same time, it drastically reduced the number of pions that were mistakenly included. In the most difficult cases, where the detector was most confused, the new method cut down the contamination from pions by a significant margin, making the data much cleaner and more reliable.

To be absolutely sure the computer was not just memorizing the simulation but actually understanding the real world, the researchers performed a special check using real experimental data. They looked for specific events where an electron or positron emitted a photon, a particle of light, as it traveled. Because pions do not emit light in this way, finding these events provided a pure sample of real electrons and positrons to test against. They also looked at events where a pion was misidentified as a positron and checked if the new filter could spot the error. The tests confirmed that the computer models worked just as well on real data as they did on simulations. The researchers found that the performance was stable across different speeds and angles, proving that the method was robust. They also noted that while the computer was excellent, there were tiny differences between the simulation and reality, so they created a simple correction factor to adjust the final numbers, ensuring that any future scientific calculations would be perfectly accurate.

This work has already been put to use in major physics studies, including measurements of how protons interact with light to produce other particles. By cleaning up the data before the scientists even began their analysis, the new algorithm allowed them to see the physics more clearly than ever before. The success of this approach shows that machine learning can be a powerful partner in high-energy physics, not just as a tool for sorting data, but as a way to see the subtle details that human-designed rules might miss. The team made their software available to the entire research community, meaning that future experiments at this facility and others can use the same smart filters to improve their own discoveries. The result is a clearer view of the subatomic world, where the line between a true electron and a confusing pion is drawn with much greater certainty.

Technical Summary: Validating a BDT-based Electron-Positron Identification Algorithm at CLAS12

Problem Statement
In the CLAS12 experiment at the Thomas Jefferson National Accelerator Facility (JLab), the identification of electrons and positrons is critical for physics measurements, particularly those relying on the detection of scattered electrons. The primary challenge arises in the Forward Detector (FD) at momenta exceeding the High Threshold Cherenkov Counter (HTCC) threshold (approximately 4.9 GeV/c). In this kinematic regime, both leptons and charged pions produce Cherenkov light, rendering the HTCC ineffective for discrimination. Consequently, the standard particle identification (PID) algorithms, which rely heavily on the HTCC and Electromagnetic Calorimeter (ECAL), suffer from significant contamination by charged pions misidentified as positrons. This background is particularly problematic for exclusive reaction analyses and cross-section measurements, where purity is essential.

Methodology
To address this contamination, the authors developed a machine-learning approach using Boosted Decision Trees (BDT). The methodology focuses on exploiting the distinct energy deposition patterns of leptons versus pions within the ECAL, specifically the Partial Sampling Fractions (SF) and the transverse size of electromagnetic (EM) showers.

Input Features: Two models were trained:
- BDT-6: Utilizes six features: partial sampling fractions (SF) and transverse shower sizes ( $m^2$ ) for the three ECAL layers (PCAL, ECin, ECout).
- BDT-9: Utilizes the six features above plus the particle's momentum, polar angle, and azimuthal angle to test for kinematic bias.
Training Data: Models were trained on simulated samples generated with GEANT4 and processed through the CLAS12 reconstruction software. The signal consisted of true electrons/positrons, and the background consisted of charged pions. The training covered momenta from 4 to 11 GeV/c to ensure coverage of the HTCC threshold region.
Validation Strategy: A rigorous validation was performed using both simulated and experimental data (liquid-hydrogen targets from Fall 2018 and Spring 2019).
- Signal Efficiency (TPR): Validated using radiative events ( $e^-p \to e^-(+)\gamma X$ ) where the angular difference between the lepton and photon ( $\Delta\theta$ ) isolates true leptons.
- Background Rejection (FPR): Validated using the exclusive process $ep \to e'\pi^+(n)$ , where the missing mass distribution reveals the presence of misidentified pions (peaking at the neutron mass).
- Simulation-Data Agreement: Input feature distributions were compared between data and simulation to derive efficiency correction functions ( $C$ ) to account for residual mismodeling.

Key Contributions

Algorithm Development: The creation of two BDT classifiers (BDT-6 and BDT-9) specifically tailored for electron/positron identification in the high-momentum regime of CLAS12.
Experimental Validation Framework: The establishment of a robust validation protocol using specific exclusive reaction channels to determine True Positive Rates (TPR) and False Positive Rates (FPR) directly from experimental data, rather than relying solely on simulation.
Efficiency Correction: The derivation of correction functions to bridge the gap between simulation-based training and experimental data performance, ensuring unbiased physics results.
Software Integration: The integration of these classifiers into the Iguana software framework, making them available for the broader CLAS collaboration.

Results
The application of the BDT models demonstrated significant improvements in background suppression while maintaining high signal retention:

Background Reduction: The models reduced charged-pion contamination by at least an order of magnitude compared to standard methods.
Performance Metrics:
- In simulated data, the models retained over 90% of leptons while significantly reducing background.
- In experimental data, the BDT-9 model generally outperformed BDT-6. For example, in the Spring 2019 dataset, the BDT-9 achieved a True Positive Rate (TPR) of 95.4% for electrons with a False Positive Rate (FPR) of 3.4%, compared to an FPR of 8.2% for BDT-6.
- For positrons in the same dataset, BDT-9 achieved a TPR of 91.6% with an FPR of 7.8%.
Stability: The classifiers showed stability across different momentum ranges (4–11 GeV/c) and magnetic field configurations (inbending and outbending), though performance varied slightly between configurations (e.g., outbending configurations showed slightly lower TPRs).
Agreement: Receiver Operating Characteristic (ROC) curves derived from experimental data were qualitatively consistent with those from simulation, validating the simulation's predictive power after applying the derived correction functions.

Significance and Claims
The paper asserts that this work provides a validated framework for advanced lepton identification at CLAS12 and future electron-beam facilities. The authors claim that their approach effectively mitigates pion contamination, a critical requirement for precision measurements such as the Timelike Compton Scattering (TCS) and near-threshold $J/\psi$ photoproduction cross-sections, which have already utilized these tools.

The significance of the work lies in its demonstration that sampling calorimeter information, when processed through machine learning, can effectively discriminate between leptons and pions in kinematic regions where traditional Cherenkov counters fail. The authors emphasize that the developed validation framework and efficiency correction methods are essential for ensuring the reliability of physics results derived from machine-learning-based identification, preventing biases in cross-section extractions. They note that these tools are now part of the collaboration-wide Iguana framework and can be extended to other datasets (e.g., deuterium and nuclear targets) and future experiments like SoLID, $\mu$ CLAS12, and ePIC.

More like this