Quantum machine learning applied for the first time to the CERN LHCb experiment

The LHCb experiment at CERN (European Centre for Nuclear Research) recently announced the first proton-proton collisions at world-record energies, with a new detector designed to cope with more demanding data acquisition conditions. The research results were recently published in the Journal of High Energy Physics under the title "Quantum machine learning for b-jet charge identification" [1].

The Data Processing and Analysis (DPA) project, led by Eduardo Rodrigues, a senior research physicist at the University of Liverpool, is a major reform of the offline analysis framework to take full advantage of the significantly increased data flow from the upgraded LHCb detector.

In this paper, the DPA team demonstrates for the first time the successful use of quantum machine learning (QML) techniques for identifying the charge of b hadron-initiated jets at the LHC. This work comes just after the start of a new data acquisition period, as part of a mid- and long-term R&D effort.

The use of machine learning techniques is ubiquitous in the analysis of the LHCb. Given the rapid development of quantum computers and quantum technologies, it is natural to start investigating whether and how quantum algorithms can be executed on such new hardware and whether use cases in LHCb particle physics can benefit from this new technology and paradigm of quantum computing. So far, QML techniques have been mainly applied to particle physics: to solve event classification and particle orbit reconstruction problems, but the team applied them for the first time to hadron jet charge identification tasks (jet classification problems).

The study was conducted based on a simulated, b hadron-initiated jet sample, and the performance of a variational quantum classifier based on two different quantum circuits was compared to deep neural network (DNN) performance on a quantum simulator. The results show that the performance of the DNN is slightly better than that of the QML algorithm, but the difference between the two is small.

Two methods of jet tagging. In the exclusive (exclusive) method, the information comes from a particle, such as a muon, whose charge is associated with a b hadron (lower jet); in the inclusive (inclusive) method, the information is extracted from the jet components (upper jet).

Two different kinds of quantum circuits. Top: Circuit representation of the Amplitude Embedding model (AEM). In blue, the variables are embedded in the amplitude of the quantum state; in red, the trainable universal rotating gate, which is optimized during the training phase; and in green, the CNOT gate, which entangles the quantum bits in a circular topology. Below: Circuit representation of the Angle Embedding model (AEM). Blue is the X-axis rotational gate for embedding variables into quantum circuits; red is the trainable universal rotational gate to be optimized during the training phase; green is the CNOT gate to entangle quantum bits with a circular topology.

The quantum algorithm performs slightly worse than the DNN, and the angular embedding circuit performs better than the amplitude embedding.

The paper shows that the QML approach is able to achieve the best performance with a smaller number of events, which helps to reduce the use of resources, which will be a key point for LHCb as the amount of data collected increases in the coming years. However, DNNs outperform QML algorithms when a large number of features are used. Improvements are expected when higher performance quantum hardware becomes available.

Research conducted in collaboration with experts has shown that quantum algorithms can allow the correlation between features to be studied. This could offer the possibility of extracting relevant information about the jet components, which would eventually improve the performance of the jet identification.

Finally, Dr. Eduardo Rodrigues said [2], "This paper demonstrates for the first time that QML can be successfully used for LHCb data analysis." The exploitation of QML in particle physics experiments is still in its infancy. As physicists gain experience with quantum computing, significant improvements in hardware and computational techniques are expected, given the global interest and investment in quantum computing.

"This work is part of the R&D activities of the LHCb Data Processing and Analysis (DPA) project and provides valuable insights into QML; it also opens up new avenues for classification problems in particle physics experiments."

Reference link：

[1]https://link.springer.com/article/10.1007/JHEP08(2022)014

[2]https://phys.org/news/2022-08-quantum-machine-lhcb.html