PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning in the Search for New Fundamental Physics

Georgia Karagiorgi, Gregor Kasieczka, Scott Kravitz, Benjamin Nachman, David Shih

arXiv:2112.03769v1hep-phhep-exphysics.data-anstat.ML

TL;DR

The review addresses how modern ML can help search for rare, subtle signals of physics beyond the Standard Model across terrestrial HEP experiments. It surveys ML methods and applications throughout experimental pipelines, finding broad progress across energy-frontier, rare-event, and neutrino searches while noting substantial computing and deployment challenges.

  • Problem

    Searches for beyond-Standard-Model physics must isolate rare, subtle signals amid large datasets, while simulations and model-specific strategies leave challenges for broad discovery.

  • Method

    The review surveys modern ML methods and applications across data acquisition, reconstruction, and final analysis in LHC, rare-event, and neutrino experiments.

  • Results

    ML is being integrated throughout HEP search pipelines, with applications including anomaly detection, weakly supervised searches, and data-driven rare-event methods.

  • Takeaways & Limitations

    Modern ML provides a complementary, less model-dependent strategy for new-physics searches, while reconstruction applications are entering workflows and other tools remain less widely deployed.

  • Takeaways & Limitations

    Fundamental-physics ML faces serious computing challenges, including training 20k neural networks for the BSM exclusion limits in Ref. [180].

Abstract

from arXiv · show

Machine learning plays a crucial role in enhancing and accelerating the search for new fundamental physics. We review the state of machine learning methods and applications for new physics searches in the context of terrestrial high energy physics experiments, including the Large Hadron Collider, rare event searches, and neutrino experiments. While machine learning has a long history in these fields, the deep learning revolution (early 2010s) has yielded a qualitative shift in terms of the scope and ambition of research. These modern machine learning developments are the focus of the present review.

I. INTRODUCTION

Modern ML is being developed to address the scale, subtlety, efficiency, and simulation challenges of terrestrial searches for physics beyond the Standard Model. This review surveys applications across event-based HEP pipelines and emphasizes that many methods remain proofs of concept while reconstruction tools are entering workflows.

  • Dark matter, dark energy, neutrino masses, and baryon asymmetry are established observations not explained by the Standard Model, motivating searches for new particles and forces.
  • Rare, subtly different new-particle signatures require sifting through immense, complex datasets, while modern deep learning can analyze large amounts of high-dimensional data for subtle patterns.
  • Efficient and accurate processing is increasingly necessary for high-throughput applications, motivating deep-learning algorithms on power-efficient hardware.
  • HEP simulations support research from experimental design through analysis but remain approximations to nature, creating a need to combine simulations with direct data information.
  • ML methods provide a complementary search strategy that is less dependent on specific models than traditional searches.
  • The review covers modern ML for terrestrial energy-frontier, rare-event, and neutrino experiments across data acquisition, reconstruction, and final analysis.These experiments use event-based data and follow similar pipelines, although recorded event structures differ.

II. MACHINE LEARNING FOUNDATIONS FOR FUNDAMENTAL PHYSICS

ML architecture is selected according to the structure and representation of physics data. Dense networks offer general flexibility, while convolutional, recurrent, recursive, graph, and permutation-invariant architectures encode different symmetries or relationships.

  • The data representation can significantly affect deep-learning design and performance, sometimes more than the algorithm or architecture.
  • Dense neural networks flatten features into vectors and are often sufficiently powerful, while physics symmetries can reduce parameters and increase performance.
  • Convolutional neural networks suit fixed-size tensors with translational invariance, such as detector images whose signatures can occur anywhere.
  • Recurrent neural networks process sequences by sharing weights across time, whereas recursive networks address hierarchical sequence structures such as jet-clustering trees.
  • Image, sequence, and tree representations can impose order where particles are physically indistinguishable, motivating architectures that process variable-length, permutation-invariant sets.
  • Graph neural networks use locality in information flow between graph nodes and network layers, providing a permutation-invariant architecture for structured data.

B. Machine Learning Tasks

ML tasks are distinguished by their targets and label availability, spanning supervised classification and regression, less-than-supervised learning, and unsupervised density-based methods. Generative approaches include GANs, VAEs, and normalizing flows.

  • Classification uses discrete finite targets and commonly employs cross-entropy, while regression uses continuous or infinite targets and commonly employs mean squared error.
  • Unsupervised learning uses no labels and typically learns the data probability density p(x), supporting generative modeling and anomaly detection.
  • Neural-network invariance and equivariance describe whether transformations of inputs leave outputs unchanged or transform outputs correspondingly.
  • Standard unsupervised deep-learning approaches include Generative Adversarial Networks, Variational Autoencoders, and Normalizing Flows.
  • GANs distinguish generated samples from data, VAEs encode and decode through a latent space, and normalizing flows use invertible functions with tractable Jacobians.
  • Supervised learning provides a label for each data point, whereas semisupervised and weakly supervised methods have labels for some data or noisy labels for all data.

III. MACHINE LEARNING SEARCHES FOR NEW PHYSICS

New-physics searches combine signal-sensitive event selection with precise Standard Model background estimation, but many current searches remain simulation-based and model-specific. Less-than-supervised ML is being developed to make searches more model-agnostic and simulation-independent, although many methods are still at the proof-of-concept stage.

  • Many direct-search methods remain at the proof-of-concept stage, while methods applied to actual data are discussed separately.
  • A new-physics search requires both selecting events enriched in signal and estimating the Standard Model background for that selection.
  • Most searches currently achieve signal sensitivity through simulation-based, model-specific strategies built around hypothesized particle parameters.
  • Less-than-supervised ML methods use data to enhance signal sensitivity and pursue more model-agnostic, simulation-independent searches.

A. Signal Model-driven / Fully Supervised Searches

Fully supervised searches use simulated signal and Standard Model samples to train classifiers that enhance sensitivity to a specified new-physics model. The resulting classifiers are applied to data to increase the presence of a potential signal.

  • A. Signal Model-driven / Fully Supervised Searches: Fully supervised classifiers are trained with simulations of signal and background to distinguish pure signal from pure background.
  • A. Signal Model-driven / Fully Supervised Searches: Earlier non-machine-learning semi-supervised searches compared many data and Standard Model simulation histograms to look for discrepancies.
  • A. Signal Model-driven / Fully Supervised Searches: Supervised searches use simulation-labeled signal and Standard Model background samples, whereas semi-supervised searches use data for either the background or signal-sensitive sample.
  • A. Signal Model-driven / Fully Supervised Searches: These classifiers are applied to data to enhance the presence of a potential signal.

1. Object Tagging

Object and event classification use machine learning to exploit complex detector representations and improve signal sensitivity. Deep learning can operate directly on low-level inputs, while analysis methods must preserve the relevant statistical structure of the search.

  • 1. Object Tagging: Jets contain tens to hundreds of particles with high- and variable-dimensional features, motivating representations as images, sequences, trees, sets, and graphs.
  • 1. Object Tagging: Boosted decision trees remain particularly common among event-level classifiers used to enhance signal sensitivity across high energy physics.
  • 1. Object Tagging: Deep learning on low-level inputs exceeded standard approaches and did not improve when hand-crafted observables were added.
  • 1. Object Tagging: Machine learning can optimize simulation-based inference, incorporate uncertainty, correct simulations, enforce feature independence, and decorrelate classifiers from resonant features.
  • 1. Object Tagging: Supervised searches test a specified signal against the Standard Model, with the likelihood ratio providing the optimal statistic when nuisance parameters are absent.
  • 1. Object Tagging: Less-than-supervised methods target low-pSM events or collective anomalies because no single event can be labeled as signal with certainty.

1. Unsupervised

Unsupervised searches seek unusual events without label information, commonly using autoencoders and other generative or clustering-based methods. Autoencoder anomaly scores can be biased by the relative complexity of anomalous and background data.

  • 1. Unsupervised: Unsupervised anomaly detection searches for data points far from the bulk of the background, commonly using autoencoders.Autoencoders encode data into a latent space and decode it back, with training based on reconstruction loss and limited transformation capacity.
  • 1. Unsupervised: Autoencoders can be biased by the relative complexity of anomalous and background data, causing some outliers to have lower loss than background events.
  • 1. Unsupervised: Variational autoencoder latent spaces can provide anomaly scores because they are optimized to follow a known background distribution.
  • 1. Unsupervised: Other investigated unsupervised approaches include support vectors, latent space Dirichlet analysis, clustering, and generative adversarial networks.

2. Weakly and Semi-Supervised

Weakly and semi-supervised searches use partial label information to improve sensitivity without fully specified signal labels, trading some of that benefit for additional assumptions. Their distinction is whether labels are noisy for all examples or noiseless for only a subset.

  • 2. Weakly and Semi-Supervised: Weakly and semi-supervised searches use label information to improve sensitivity for BSM particles, at the cost of additional assumptions.
  • 2. Weakly and Semi-Supervised: Weak supervision uses noisy labels for every example, whereas semi-supervision uses noiseless labels for only a subset of training examples.
  • 2. Weakly and Semi-Supervised: Weak supervision commonly separates data into mixtures A and B containing the same signal and background classes but with different unknown signal fractions.
  • 2. Weakly and Semi-Supervised: The mixture-based weakly supervised approach works only when the signal and background classes are statistically identical across A and B.Otherwise, the classifier can learn differences between the mixtures that are unrelated to the signal.
  • 2. Weakly and Semi-Supervised: The review organizes applications across energy, neutrino, and rare-event frontiers and across data acquisition, reconstruction, and final analysis.

A. Energy Frontier

At the LHC, machine learning supports the full search pipeline, from high-rate event selection and reconstruction to object tagging, background estimation, and final discriminants. Supervised methods dominate current event-level analyses, while weakly supervised approaches are beginning to identify anomalous regions with limited data results.

  • Pipeline-wide applications: LHC searches use machine learning across data acquisition, reconstruction, calibration, object tagging, background estimation, and final analysis discriminants.Applications span detector-level processing and analysis-level selection of potential new-particle signatures.
  • Data acquisition: 40 MHz collisions and O(MB) event sizes require online processing that reduces the event rate to 1 kHz by discarding more than 99.99% of events.ATLAS and CMS use two-stage processing, including a Level-1 trigger, to meet the data-rate constraint.
  • Object tagging: A fully connected DNN b-tagging algorithm achieves a 1/390 light-flavor false-positive rate while retaining 70% true b-jet efficiency.The example illustrates deployed ML-based object tagging for enriching samples with relevant physics processes.
  • Background estimation: Machine learning supports robust background estimation through decorrelation, high-dimensional reweighting, and mitigation of data/simulation differences.These strategies are used to enable resonance searches and improve agreement between control-region data and background models.
  • Final analysis: Supervised learning remains the most common event-level approach, whereas weakly supervised dijet analysis extended inclusive-search sensitivity by identifying anomalous phase-space regions.The weakly supervised result used only two features, but complete data results remain limited by analysis time.

B. Neutrino Experiments

Neutrino experiments apply machine learning throughout their data pipelines to study neutrino masses and possible departures from the three-neutrino paradigm. Detector geometry and data structure shape the architectures, including CNNs for image-like detectors and graph networks for irregular configurations.

  • Scientific scope: Neutrino searches investigate neutrino masses and deviations from the three-neutrino paradigm, including sterile neutrinos, new mediators, extra dimensions, and non-standard interactions.Classifying the interacting neutrino flavor is a key requirement for these analyses.
  • Pipeline-wide applications: Deep learning applications span data acquisition, reconstruction, and final analysis, with reported improvements in performance and physics reach over traditional methods.The review describes growing use across the full neutrino data pipeline.
  • Image recognition: LArTPCs provide stereoscopic image streams suited to computer-vision methods, although current deep-learning applications mainly focus on reconstruction and final analysis.DUNE, MicroBooNE, and SBN use this detector technology.
  • Image recognition: CNNs and sparse CNNs support neutrino interaction classification, kinematic reconstruction, semantic segmentation, and cosmic-background removal in LArTPC experiments.These methods target both event-level classification and pixel-level reconstruction tasks.
  • Geometry-aware methods: Detector geometry motivates specialized architectures, including Siamese CNNs for NOνA projections, spherical CNNs for KamLAND-Zen, and graph neural networks for IceCube’s irregular configuration.IceCube reports significant physics-performance improvement over traditional algorithms using graph networks.
  • Data acquisition: DUNE far detectors are expected to generate several terabytes per second of raw data while requiring 100% live-time for sensitivity to stochastic signals.The scale creates data-acquisition demands approaching those of collider experiments.

C. Rare Event Searches

Rare event searches use machine learning for background rejection, waveform and image analysis, and event reconstruction, but their few-event sensitivity makes simulation mismodeling especially consequential. Across experiments, data representation can matter more than the specific algorithm or architecture.

  • Scope and challenge: Dark-matter and neutrinoless double-beta-decay experiments seek BSM processes but use machine learning less prevalently than collider and neutrino experiments.The passage associates this partly with sensitivity to mismodeling at the few-event level.
  • Shallow and time-series methods: BDTs, neural networks, and CNNs are used for rare-background removal, pulse-parameter extraction, pulse-shape discrimination, and waveform summarization.Applications include xenon, solid-state, bubble-chamber, and germanium detectors.
  • Image recognition: 2D CNNs exploit spatial or spatiotemporal hit patterns for topological signal-background discrimination in xenon time-projection chambers.The approach has been implemented in experiments including NEXT, nEXO, and PandaX-III.
  • Less-than-supervised searches: Less-than-supervised methods reduce reliance on simulations through data-driven, semi-supervised, and unsupervised training strategies.Examples include pseudo-label iteration in PICO, physics-invariant augmentation in NEXT, and clustering methods in CRESST.
  • Representation and architecture: Data representation often affects results more than algorithm or architecture: PICO improved discrimination with Fourier-space features, while EXO-200 saw minimal CNN gains over a feature-engineered BDT.DarkSide-20k also improved with a fully connected network over a CNN, while latent-space classification performed well in germanium.
  • Open challenges: Future progress requires extracting physical insight from unsupervised methods, quantifying uncertainties, and constructing training sets that reduce domain discrepancies.These priorities are stated as necessary for fully exploiting deep learning in rare event searches.

V. OUTLOOK AND CHALLENGES

Machine learning is increasingly important for searching high-dimensional data for rare, subtle deviations from the Standard Model and for reducing model dependence. Its deployment remains constrained by domain-specific challenges, computing demands, data-dependent selections, and the need for new analysis and reinterpretation practices.

  • Outlook: New particles may be discoverable with existing and near-future experiments, motivating exploration of data in its natural high-dimensionality and reduced model dependence.The review describes rapid development of ML applications in data processing and analysis.
  • Domain challenges: Fundamental-physics ML must detect ultra-rare, subtle deviations where statistical evidence emerges from many examples rather than a uniquely anomalous event.These characteristics create challenges not necessarily solved by approaches developed for industry.
  • Computing and reproducibility: The BSM exclusion limits in one search required training 20k neural networks, illustrating the computing demands of less-than-supervised analyses.The selection changes when a different amount of signal is injected, requiring repeated training.
  • Analysis infrastructure: Less-than-supervised searches require new approaches to data preservation and analysis reinterpretation because event selection depends on the injected signal amount.This dependence complicates reuse of analyses under different signal hypotheses.
  • Broader applications: ML is also expanding into indirect searches, including data corrections with BSM effects, effective field theory analyses, and reinterpretation.The review notes related opportunities beyond direct searches for new particles.
Loading 2112.03769v1…