Searching for Exotic Particles in High-Energy Physics with Deep Learning

Abstract: Collisions at high-energy particle colliders are a traditionally fruitful source of exotic particle discoveries. Finding these rare particles requires solving difficult signal-versus-background classification problems, hence machine learning approaches are often used. Standard approaches have relied on `shallow' machine learning models that have a limited capacity to learn complex non-linear functions of the inputs, and rely on a pain-staking search through manually constructed non-linear features. Progress on this problem has slowed, as a variety of techniques have shown equivalent performance. Recent advances in the field of deep learning make it possible to learn more complex functions and better discriminate between signal and background classes. Using benchmark datasets, we show that deep learning methods need no manually constructed inputs and yet improve the classification metric by as much as 8\% over the best current approaches. This demonstrates that deep learning approaches can improve the power of collider searches for exotic particles.

• The paper demonstrates that deep neural networks capture complex signal-background patterns without manual feature engineering, enhancing classification accuracy by up to 8%.
• The paper shows that deep learning outperforms traditional shallow models in identifying rare particle signals from high-dimensional collider data.
• The paper validates its approach using Higgs boson and supersymmetry benchmarks, highlighting deep learning’s promise for future high-energy physics research.

Overview of "Searching for Exotic Particles in High-Energy Physics with Deep Learning"

The paper entitled "Searching for Exotic Particles in High-Energy Physics with Deep Learning," authored by P. Baldi, P. Sadowski, and D. Whiteson, explores the application of deep learning (DL) methods to classify signal and background data in high-energy physics (HEP), specifically within the domain of exotic particle searches. The authors argue that existing shallow ML approaches, while useful, may not fully capture complex non-linear relationships present in collision data. They demonstrate that deep neural networks (DNNs) exhibit improved performance over traditional methods, providing an avenue for enhanced anomaly detection in particle collider environments.

Context and Problem Setting

The discovery of exotic particles at high-energy colliders relies on effectively distinguishing rare signals from prevalent background events within high-dimensional datasets. Traditional ML approaches have used shallow models requiring manual feature engineering, which is labor-intensive and potentially sub-optimal. The need for robust methods that automatically learn from raw data without significant manual intervention is thus pivotal. This study introduces deep learning to address these limitations by autonomously identifying complex signal-background dichotomies, utilizing benchmark datasets that simulate conditions at particle colliders like the Large Hadron Collider (LHC).

Methodology and Experiments

The researchers employed DNNs with architectures optimized through hyper-parameter tuning and compared their performance against established methodologies, such as shallow neural networks (NN) and boosted decision trees (BDT). Leveraging benchmark cases involving Higgs boson and supersymmetric (SUSY) particles, they demonstrated the DNN’s capacity to classify collision events without reliance on manually contrived features.

• Higgs Boson Benchmark: The task aimed to distinguish signal processes from background, simulating events involving heavy electrically-neutral Higgs bosons. The study utilized specific high-level and low-level features from collision data to train the ML models.
• Supersymmetry Benchmark: Addressing the SUSY particle detection, the challenge was to differentiate between processes producing novel supersymmetric particles and those resulting from the production of $W$ boson pairs, using observable and inferred data from collision events.

Both scenarios deployed DNNs that surpassed shallow models in performance as measured by the Area Under the ROC Curve (AUC) metric and improved discovery significance, indicative of enhanced sensitivity to signal presence over background noise.

Key Results

The findings noted a performance increase, with DNNs achieving up to 8% higher classification accuracy (AUC) compared to shallow NNs or BDTs. The authors emphasized how deep architectures learned more effective feature combinations, thus capturing untapped discriminative potential from raw input data that were not fully realized by traditional ML methods with manual feature augmentation.

Implications and Future Directions

The demonstrated ability of DNNs to autonomously derive complex feature representations without prior manual engineering signifies a substantial shift in the analytical toolkit available for exotic particle searches in HEP. The researchers suggest that the procedural advancements seen in deep learning can extend beyond benchmark studies, potentially impacting more nuanced challenges such as multiple background differentiation or lower-level data processing tasks.

Looking toward future advancements in AI, the successful integration of deep learning in high-energy physics could influence broader scientific domains where similar high-dimensional, non-linear problems prevail, enhancing methodologies for anomaly detection and complex pattern analysis.

The paper advances the field by situating deep learning as a superior analytical method, poised to transform signal extraction and anomaly recognition tasks central to high-energy physics, paving the way for future investigations into more intricate particle phenomena and firmly establishing its relevance in scientific discovery processes.