QCD-Aware Recursive Neural Networks for Jet Physics (1702.00748v2)

Abstract: Recent progress in applying machine learning for jet physics has been built upon an analogy between calorimeters and images. In this work, we present a novel class of recursive neural networks built instead upon an analogy between QCD and natural languages. In the analogy, four-momenta are like words and the clustering history of sequential recombination jet algorithms is like the parsing of a sentence. Our approach works directly with the four-momenta of a variable-length set of particles, and the jet-based tree structure varies on an event-by-event basis. Our experiments highlight the flexibility of our method for building task-specific jet embeddings and show that recursive architectures are significantly more accurate and data efficient than previous image-based networks. We extend the analogy from individual jets (sentences) to full events (paragraphs), and show for the first time an event-level classifier operating on all the stable particles produced in an LHC event.

• The paper presents a novel method that treats jet particle momenta like words in a sentence for enhanced jet analysis.
• It leverages a recursive neural network architecture grounded in QCD principles to preserve critical jet substructure details.
• Empirical results show superior accuracy and data efficiency in jet and event classification compared to traditional image-based models.

Overview of QCD-Aware Recursive Neural Networks for Jet Physics

The paper presents an innovative approach to jet physics by developing a class of recursive neural networks that draws an analogy between quantum chromodynamics (QCD) and natural language processing. The researchers propose a novel methodology wherein the four-momenta of particles are treated similarly to words in sentences, and jet clustering history is akin to sentence parsing. This approach offers a new paradigm distinct from existing methods that equate jet images to pixels in calorimeters.

Core Methodology and Results

The research introduces a recursive neural network (RNN) architecture that processes jets by leveraging their substructure directly from QCD principles rather than through the loss-prone transformation into images. Each jet is represented as a variable-length sequence of four-momenta organized into an adaptive tree structure, where the variation is contingent on the particular event. This methodology captures subtleties in jet physics overlooked by conventional image-based models.

The paper provides empirical evidence that recursive architectures, grounded in a QCD-native context, deliver superior accuracy and data efficiency relative to those reliant on image projection. Constructing jet embeddings based on variable-length sets of particles without projecting into fixed grids preserves critical information lost in current techniques, particularly those using large neural networks such as MaxOut with extensive parameters requiring substantial data for training.

Experimental Findings

Key results are drawn from jet and event-level classification tasks. Recursive neural networks, particularly when topology reflects physically motivated structures like $k_t$ or Cambridge-Aachen clustering, exhibit higher performance metrics than simpler sequence-based or random structures. For example, ROC AUC scores and signal efficiency metrics demonstrate quantifiable superiority in discriminating signal from background noise in particle measurements compared to calorimeter simulations.

Extending the analogy to events, full event embeddings are implemented, providing the capability to embed entire Large Hadron Collider (LHC) events as sequences of jets. This hierarchical approach encapsulates the information of all detectable particles, suggesting a robust framework for handling the intricacies of hadron-level event classification tasks.

Implications and Future Directions

This QCD-aware approach holds potential for substantial advancements in particle physics, particularly in contexts requiring precise jet structure analysis, such as boosted particle identification. The recursive paradigm not only enhances performance measures but also aligns closely with domain-specific physical realities, ensuring potentially greater robustness to variations.

Speculating on future capabilities, adapting the recursive structures to be more adaptive, learning the optimal topology for given classification tasks, or integrating advanced generative models into these architectures could further leverage machine learning's strengths in handling high-dimensional and complex particle data. Additionally, ensuring infrared and collinear safety within these frameworks could formalize theoretical guarantees vital for adopting deep learning models in high-energy physics.

Conclusion

The paper represents a meaningful contribution to jet physics by successfully marrying deep learning's recursive methodology with QCD principles, offering a more nuanced and efficient analysis of jets. As machine learning continues to interface closely with experimental physics, such methodologies reinforce the potential harmonization between data-driven approaches and rich domain expertise. This QCD-aware recursive network signifies a step towards sophisticated, adaptive models in the ever-evolving landscape of particle physics.