Distribution of the anomaly rating for information and five standard BSM models based on the AE. Credit: Physical Review Letters (2024). DOI: 10.1103/PhysRevLett.132.081801
Researchers utilized a neural system, a brain-inspired machine learning algorithm, to sift through extensive amounts of particle motion data. This vast and expanding dataset of motion information is being scrutinized by particle physicists in the quest for traces of unidentified particles. Specifically, they are on the lookout for particles not accounted for in the Standard Model of particle physics, the current framework of the universe that scientists suspect may be incomplete.
As part of the ATLAS initiative, scientists at the U.S. Department of Energy’s Argonne National Laboratory and their partners have recently examined substantial volumes of ATLAS data through a machine learning method known as anomaly detection. This marks the first time that data from an accelerator experiment has undergone such analysis. This approach holds the potential to enhance the efficiency of the collaboration’s pursuit of novel discoveries. The partnership involves experts from 172 research organizations.
To identify anomalies or peculiarities in the data, the team employed a brain-inspired form of machine learning system referred to as a neural network. This methodology diverges from traditional approaches to scientific exploration, as it operates free from the biases of experts and operates independently of them.
ATLAS researchers have traditionally relied on established theoretical models to guide their research and steer their analyses toward the most promising avenues for breakthroughs. This often entails conducting intricate computer simulations to predict the expected outcomes of specific collision data based on the Standard Model.
These predictions derived from the Standard Model are then juxtaposed by researchers against actual ATLAS data, as well as compared to forecasts generated by innovative physics models that attempt to elucidate phenomena like dark matter and other enigmatic aspects of the universe.
Despite the trillions of recorded events at ATLAS, no deviations from the Standard Model have been observed thus far. Since the discovery of the Higgs boson in 2012, the ATLAS experiment has yet to unearth any new particles.
Sergei Chekanov, a scientist in Argonne’s High Energy Physics division and the primary author of the study, asserts that anomaly monitoring represents a distinctive approach to conducting this research. The primary aim is to identify peculiar patterns in the data that may appear entirely unfamiliar and unrelated to existing theoretical frameworks, rather than focusing solely on specific deviations.
In order to conduct this form of analysis, the researchers represented each particle interaction within the data as an image resembling a QR code. Subsequently, the team trained their neural network by exposing it to 1% of these images.
One of the eight activities that contributed to the most significant deviation from the Standard Model predictions that the neural network in this study discovered, displays an ATLAS event display. Credit: Lhc
Approximately 2 million interconnected networks, akin to brain cells, constitute the neural network. Without external guidance, it autonomously identified and retained connections between images representing Standard Model interactions. In essence, it acquired the ability to recognize common occurrences consistent with the predictions of the Standard Model.
Following the training phase, the scientists inputted the remaining 99% of the images into the neural network to detect any discrepancies. When presented with an image as input and tasked with reconstructing it based on its comprehensive knowledge of the dataset, the neural network may struggle if encountering novel or unusual patterns.
Chekanov explained that when faced with unfamiliar or anomalous data, the neural network encounters challenges and may struggle to accurately reconstruct the image. Significant disparities between the input image and the reconstructed output signal potential areas of interest for further investigation.
The neural system scrutinized approximately 160 million events from LHC Run-2 data collected between 2015 and 2018 using computational resources at Argonne’s Laboratory Computing Resource Center.
One anomaly identified by the neural network, which experts believe warrants further exploration despite not exhibiting definitive signs of new physics within this dataset, involves an extraordinary electron decay event occurring at a rate of about 4. This event involves the production of a mu (a fundamental particle) in combination with a cascade of other particles at 8 teraelectronvolts, contradicting the neural network’s understanding of Standard Model interactions.
“We need to delve deeper into this discovery,” stated Chekanov. “While it may represent a quantitative deviation, there is a possibility that this decay could signify the presence of a previously unknown particle.”
The team plans to apply this methodology during the LHC Run-3 phase, which commenced in 2022. ATLAS researchers aim to explore the potential of machine learning and anomaly detection as tools for uncovering uncharted territories within the realm of particle physics.
The article detailing this research has been published in the Physical Review Letters journal.
More details: G. Aad and colleagues, Physical Review Letters (2024), Use of Unsupervised Machine Learning for Anomaly Detection at s=13 TeV with the ATLAS Detector, Search for New Phenomena in Two-Body Invariant Mass Distributions. DOI: 10.1103/PhysRevLett.132.081801
Citation: Machine learning could help reveal undiscovered particles within data from the Large Hadron Collider (2024, April 15) retrieved 15 April 2024 from https://phys.org/news/2024-04-machine-reveal-undiscovered-particles-large.xml
This report is subject to rights. No part may be reproduced without the prior written consent, aside from any good struggling made for the purpose of private investigation or research. The information is only provided for informational purposes.
