Variational Autoencoders for New Physics Mining at the Large Hadron Collider

Olmo Cerri; Thong Q. Nguyen; Maurizio Pierini; Maria Spiropulu; Jean-Roch Vlimant

Source-linked AI summary

Variational Autoencoders for New Physics Mining at the Large Hadron Collider

Olmo Cerri, Thong Q. Nguyen, Maurizio Pierini, Maria Spiropulu, Jean-Roch Vlimant

arXiv:1811.10276v3Nov 26, 2018hep-excs.LGhep-ph

TL;DR

LHC searches can miss new physics because conventional analyses depend on specified models and trigger selections discard most collisions. The paper trains VAEs on known SM processes and applies a reconstruction-loss threshold to select anomalous events, finding that challenging BSM models can yield a high-purity event stream while exposing model-dependence and efficiency trade-offs.

Problem
LHC BSM searches typically test specific models, while trigger selection can reject new-physics events that do not match those assumptions.
Method
The paper trains variational autoencoders on a known-SM cocktail and uses reconstruction loss with a one-sided threshold to select anomalous events.
Results
The strategy can select datasets enriched with events from challenging BSM models in an LHC trigger setting.
Takeaways & Limitations
An anomalous-event stream could be scrutinized or cataloged, with repeated topologies informing future BSM models and supervised searches.
Takeaways & Limitations
Using 21 high-level features leaves residual model dependence, and the high-purity stream has relatively small signal efficiency and a strong dataset-definition bias.

Abstract

from arXiv · show

Using variational autoencoders trained on known physics processes, we develop a one-sided threshold test to isolate previously unseen processes as outlier events. Since the autoencoder training does not depend on any specific new physics signature, the proposed procedure doesn't make specific assumptions on the nature of new physics. An event selection based on this algorithm would be complementary to classic LHC searches, typically based on model-dependent hypothesis testing. Such an algorithm would deliver a list of anomalous events, that the experimental collaborations could further scrutinize and even release as a catalog, similarly to what is typically done in other scientific domains. Event topologies repeating in this dataset could inspire new-physics model building and new experimental searches. Running in the trigger system of the LHC experiments, such an application could identify anomalous events that would be otherwise lost, extending the scientific reach of the LHC.

1 Introduction

LHC searches for beyond-Standard-Model physics are typically model dependent, motivating an unsupervised VAE trigger strategy to identify anomalous events without a specific new-physics signature.

Motivation: Model-dependent hypothesis tests may miss future BSM scenarios outside the models considered, while LHC trigger limits can reject potentially interesting events before storage.The LHC produces 40 million collisions per second but stores only about 1000 events per second.
Proposed strategy: The paper proposes training an unsupervised algorithm on known SM processes to identify BSM events as anomalies and store them for expert scrutiny.The resulting events could also be released as an open-access catalog.
VAE trigger: A VAE replaces a model-dependent trigger selection by compressing single-lepton events into a stochastic latent space and reconstructing their input distributions.The procedure is presented for a single-lepton stream but could extend to other data streams.
Threshold selection: A reconstruction-loss threshold defines an anomalous-event stream, calibrated to collect approximately 1000 SM events per month under typical operating assumptions.Events above the threshold are treated as potential anomalies.
Benchmarking: The study evaluates BSM production cross sections corresponding to 100 selected BSM events per month or a signal yield about one-third of the SM yield.The benchmark considers low-mass BSM resonances decaying to one or more leptons.

2 Related work

Prior model-independent searches compare many simulated distributions for excesses, while this work uses deep-learning anomaly detection in the trigger to retain events that conventional online selection could discard.

Existing searches: Earlier model-independent searches compare large sets of binned data distributions with Monte Carlo predictions to find unusually large deviations.The effectiveness of this strategy for establishing a discovery has been debated.
Distinction from prior work: This strategy aims to process events that online selection might discard by running anomaly detection as part of the trigger process.That trigger-level focus distinguishes it from approaches applied only after online selection.
Deep-learning methods: The paper uses variational autoencoders based on high-level features as a baseline for searches requiring minimal or no assumptions about the new-physics scenario.Related work has also explored autoencoders for detector monitoring, event generation, and anomalous-jet tagging.

3 Data samples

The study builds a simulated single-lepton dataset under Run-II-like conditions, models four dominant SM processes and several BSM benchmarks, and mixes the SM samples into a training cocktail.

Dataset construction: The dataset uses PYTHIA8 proton-proton collisions at 13 TeV with pileup of 20, followed by DELPHES simulation with the CMS HL-LHC detector card.These conditions loosely correspond to LHC operating conditions in 2016.
SM samples: The SM cocktail contains the four processes with the highest production cross sections, with samples scaled to the lowest-statistics QCD sample.The cocktail trains autoencoders and tunes the anomaly threshold using production cross sections and selection efficiencies.
Event selection: Events are selected with one reconstructed electron or muon having pT > 23 GeV and loose isolation Iso < 0.45.If multiple reconstructed leptons are present, the highest-pT lepton is used.
Event representation: The event representation uses 21 high-level quantities covering lepton, isolation, jet, missing-momentum, transverse-mass, and multiplicity information.The feature list is designed to represent the main SM-process information rather than a specific BSM scenario.
BSM benchmarks: BSM benchmarks include a leptoquark, neutral and charged scalars, a scalar decaying to tau pairs, and QCD multijet production.For each BSM scenario, direct PYTHIA8 production mechanisms, including associated jet production, are considered.

4 Model description

The model uses a variational autoencoder trained on Standard Model events represented by 21 high-level features, learning probabilistic latent and reconstructed representations for anomaly detection. Its loss combines reconstruction likelihood with a weighted KL divergence, while feature-specific output distributions improve modeling of distribution cores and tails.

4.1 Autoencoders: The VAE is trained on a Standard Model cocktail using 21 high-level features, although this representation leaves residual model dependence in anomaly detection.The selected features are not tailored to specific BSM models, but performance may not generalize equally across scenarios.
4.1 Autoencoders: VAEs model probability distributions in latent and original spaces, unlike plain autoencoders, which produce point estimates.This provides both a best-point estimate and an estimate of associated statistical noise.
4.1 Autoencoders: The architecture encodes 21 inputs into a four-dimensional Gaussian latent space, samples latent variables, and decodes them into predicted input distributions.The encoder and decoder use fully connected layers, with four-node parameter layers for latent means and widths.
4.1 Autoencoders: The total loss combines reconstruction negative log-likelihood with a weighted KL divergence between the latent distribution and its prior.The weight is fixed to β = 0.3, and the prior is a four-dimensional Gaussian with learnable means and diagonal covariance terms.
4.1 Autoencoders: Feature-specific probability functions replace a standard MSE loss to better describe the observed distributions, especially their cores and tails.The model uses clipped log-normal, truncated Gaussian, discrete truncated Gaussian, binomial, and Poisson forms for different feature classes.
4.1 Autoencoders: The final performance depends on the chosen conditional feature distributions and latent prior, motivating future studies with improved choices and real collision data.Learning the prior from data was explored but provided no practical anomaly-detection advantage in this study.

5 Results with VAE

The VAE identifies anomalous events using a reconstruction-loss threshold calibrated to retain about 5.4 · 10^-6 of SM events, while probing multiple BSM benchmarks with model-independent sensitivity. Its selection improves signal-over-background relative to a traditional single-lepton trigger and captures events beyond simple feature-tail outliers.

Anomaly selection: The VAE threshold retains 5.4 · 10^-6 of SM events, corresponding to about 1000 selected events per month under the assumed LHC conditions.The threshold is applied to the reconstruction loss, and the selected sample’s background composition is reported by process.
Anomaly selection: The VAE’s reconstruction loss provides stronger discrimination than the KL-divergence component, whose contribution to the total loss is negligible.Anomalies are defined from the right tail of the expected reconstruction-loss distribution, although limited training statistics can leave the threshold p-value uncalibrated.
Event characteristics: The VAE-selected anomalous events span the input-feature ranges rather than clustering only in distribution tails.This behavior is shown for both SM events and the A →4ℓ benchmark, indicating selection beyond simple feature-outlier triggers.
Benchmark comparison: The VAE probes four BSM scenarios simultaneously with comparable performance, while supervised BDTs outperform it on their targeted benchmarks.The result illustrates a trade-off between precision and model independence and complements supervised searches.
Benchmark comparison: The VAE reaches BSM cross-section sensitivities comparable to existing exclusion bounds for the studied mass ranges, including a low-mass leptoquark example limited by trigger coverage.Table 4 reports efficiencies and cross sections corresponding to 100 selected events per month and to a signal-over-background ratio of 1/3.
Purity comparison: The VAE selection achieves a signal-over-background ratio about two orders of magnitude larger than a traditional inclusive single-lepton trigger.The comparison is based on the ratio of selection efficiencies for SM and BSM events.

6 How to deploy a VAE for BSM detection

The proposed deployment uses VAEs in dedicated trigger streams to retain a small daily sample of anomalous events for later scrutiny. Training on data is presented as feasible, with contamination tests and threshold monitoring addressing practical deployment conditions.

Trigger deployment: A VAE trigger could isolate a limited number of anomalous events daily into a dedicated dataset for visual inspection or further model-independent analysis.The proposed stream is intended for follow-up analysis rather than early discovery alone.
Trigger deployment: The anomaly threshold must be monitored on real data and adjusted when necessary during HLT operations.This operational step follows the proposed offline training procedure.
Train-on-data strategy: Injecting about 700 A →4ℓ events into the training sample tests a train-on-data strategy at a 7.1 pb signal cross section and 100 pb^-1 luminosity.The injected contamination corresponds to the value at which the VAE would select 100 A →4ℓ events in one month.
Train-on-data strategy: Training on collected data could simplify deployment by incorporating data systematics automatically and improving robustness to effects affecting MC-based training.The paper identifies energy scale and efficiency as examples of such systematic effects.

7 Conclusions and outlook

The proposed VAE strategy targets generality rather than optimal sensitivity to any particular BSM model, producing anomaly samples for expert scrutiny and future search development. Its high-purity operating point trades away signal efficiency and introduces dataset bias, while broader deployment could examine more LHC events.

7 Conclusions and outlook: Events from challenging BSM models could be isolated in a sample of about 30% purity, with roughly 43 events selected per day.The quoted operating point targets potentially interesting events rather than enhanced BSM signal-selection efficiency.
7 Conclusions and outlook: The purity target incurs relatively small signal efficiency and strong dataset-definition bias, making the selected events difficult to use in traditional supervised searches.This limitation follows from the application’s focus on a high-purity anomalous-event sample rather than maximal signal efficiency.
7 Conclusions and outlook: The resulting anomalous-event list could be scrutinized or released as a catalog, with recurring patterns motivating BSM scenarios and future supervised searches.The catalog is intended as a source of events for expert follow-up and subsequent model building.
7 Conclusions and outlook: The VAE’s strength is model independence and generalization to unforeseen BSM scenarios, not guaranteed optimal sensitivity for any individual hypothesis.A supervised BDT can discriminate a given BSM hypothesis better, but the VAE trades discrimination power for broader applicability.
7 Conclusions and outlook: The study presents the strategy as a possible way to extend the physics reach of current and future LHC operations.This is stated as an expected application-level benefit rather than a demonstrated discovery result.
7 Conclusions and outlook: Although demonstrated on a single-lepton stream, the approach could be generalized to other L1-selected streams and scrutinize the approximately 100 Hz rate entering the HLT.The L1 selection remains a potentially dangerous bias, while HLT deployment could access more events than typical offline studies.

A Comparison with Auto-Encoder

A plain autoencoder provides competitive performance for some SM acceptance rates, but the VAE generally performs better at the study’s chosen threshold. The VAE’s advantage reaches two orders of magnitude for one benchmark model, with one exception favoring the AE.

A Comparison with Auto-Encoder: At ϵSM = 5.4 · 10^-6, the VAE usually outperforms the plain AE on BSM-model efficiency.The comparison uses an AE with four latent neurons and MSE loss against the VAE architecture used in the study.
A Comparison with Auto-Encoder: For the A →4ℓ model, the VAE improves efficiency by as much as two orders of magnitude over the AE.This is the largest improvement reported in the supplied comparison.
A Comparison with Auto-Encoder: For h± →τν, the AE provides 30% larger efficiency than the VAE.This is the exception to the general VAE advantage at the selected SM acceptance rate.