Quantum Machine Learning Framework Achieves 95% Confidence in Double Higgs Searches with 10% and 50% Uncertainty Control

The search for rare processes at the Large Hadron Collider receives a significant boost from a new framework developed by Marwan Ait Haddou of Hassan II University of Casablanca, Mohamed Belfkir and Salah Eddine El Harrauss of United Arab Emirates University, and colleagues. Their work introduces a Hybrid Quantum Machine Learning (HyQML) approach designed to enhance the sensitivity of searches for pairs of Higgs bosons, elusive particles central to our understanding of mass. By combining the power of quantum circuits with classical neural networks, the team achieves a two-fold improvement in performance over existing methods, allowing for more precise measurements of the Higgs boson’s interactions and tighter constraints on fundamental parameters governing the universe. This advancement promises to unlock new insights into the Higgs boson’s self-coupling and its interactions with other particles, potentially revealing physics beyond the Standard Model.

The researchers aim to enhance the sensitivity of searches for new physics through improved signal-background discrimination, addressing a significant challenge in particle physics where identifying rare events requires sophisticated data analysis techniques. Scientists utilized simulated data mirroring conditions at the LHC, representing both the desired Higgs pair production signal and background processes. They employed deep neural networks as a baseline and innovative hybrid quantum-classical models integrating variational quantum circuits with classical neural networks, leveraging the strengths of both approaches.

A key innovation involves a meta-learning technique, addressing the vanishing gradient issue that hinders quantum machine learning training. The models analyze high-level kinematic features to distinguish between signal and background. Results demonstrate that the hybrid quantum-classical models outperform classical deep neural networks in separating signal from background, stemming from the ability of the meta-learning technique to effectively train the quantum circuits. This enhanced discrimination power translates into a potential increase in the sensitivity of searches for new physics beyond the Standard Model, allowing scientists to probe more deeply into the fundamental laws of nature and offering a pathway towards more precise measurements. This research addresses the need for methodological improvements in collider physics, where current searches leave room for increased sensitivity to rare particle interactions. The core of this work lies in a hybrid architecture combining the strengths of both quantum and classical machine learning.

Researchers mapped event-level features into a quantum feature space using parameterized quantum circuits, allowing for the encoding of data into quantum states and the execution of transformations that capture complex correlations. Simultaneously, a classical neural network was integrated to retain optimization stability and scalability, crucial for handling large datasets. This combination allows the model to learn complex patterns while remaining computationally manageable. To assess the model’s performance, the HyQML framework was trained to discriminate between signal and background events. The team rigorously compared the HyQML model against a state-of-the-art XGBoost model and a purely classical implementation, demonstrating a factor of two improvement in performance.

This translates to tighter constraints on the non-resonant double Higgs boson production cross-section and improved estimations of the Higgs boson self-coupling and quartic vector-boson-Higgs coupling, highlighting the potential of quantum-enhanced learning for collider physics applications. The improved performance allows scientists to more precisely measure the properties of the Higgs boson and search for deviations from the Standard Model of particle physics. The team successfully combined parameterized quantum circuits with a classical neural network, embedding event-level features into a feature space while maintaining optimization stability. Results demonstrate that the HyQML model outperforms both a state-of-the-art XGBoost model and a purely quantum implementation, achieving a factor of two improvement in performance. Analysis of the quantum circuit’s internal representations revealed a progressive separation of signal and background distributions during training.

Initial projections showed overlapping distributions, but later epochs demonstrated clear clustering, indicating the model’s ability to learn discriminative patterns. To assess statistical sensitivity, the team optimized score regions to maximize expected significance. The HyQML model achieved an expected 95% confidence level upper limit on the non-resonant double Higgs boson production cross-section under background normalization uncertainties, revealing improved constraints on the Higgs boson self-coupling and quartic vector-boson-Higgs coupling compared to purely classical and quantum models. The improved sensitivity allows scientists to probe the Higgs boson’s properties with greater precision and search for subtle deviations from the Standard Model. The team successfully combined parameterized quantum circuits with a classical neural network, embedding event-level features into a feature space while maintaining optimization stability. The HyQML model demonstrably outperforms both a state-of-the-art XGBoost model and a purely classical implementation, achieving a factor of two improvement in performance, translating to tighter constraints on the non-resonant double Higgs boson production cross-section.