Universal New Physics Latent Space (2407.20315v1)

Abstract: We develop a machine learning method for mapping data originating from both Standard Model processes and various theories beyond the Standard Model into a unified representation (latent) space while conserving information about the relationship between the underlying theories. We apply our method to three examples of new physics at the LHC of increasing complexity, showing that models can be clustered according to their LHC phenomenology: different models are mapped to distinct regions in latent space, while indistinguishable models are mapped to the same region. This opens interesting new avenues on several fronts, such as model discrimination, selection of representative benchmark scenarios, and identifying gaps in the coverage of model space.

• The paper develops a machine learning approach that embeds BSM theories into a latent space, clearly distinguishing between phenomenological models using contrastive loss.
• The method clusters models based on key parameters like gluino-neutralino mass gaps and jet observables, enhancing benchmark selection and experimental reinterpretation.
• The approach enables model-agnostic anomaly detection and aids in solving inverse problems, paving the way for future advances in high-energy collider research.

Universal New Physics Latent Space: A Critical Examination

The paper under review presents a novel approach centered on mapping high-dimensional Beyond the Standard Model (BSM) theories into a low-dimensional latent space using a ML methodology, with the aim of distinguishing between different phenomenological signatures observed at the Large Hadron Collider (LHC). The work by Anna Hallin, Gregor Kasieczka, Sabine Kraml, André Lessa, Louis Moureaux, Tore von Schwartz, and David Shih represents a significant exercise in applying ML techniques to facilitate new physics explorations.

Methodology Overview

The core of the methodology is an encoder network, specifically a fully connected multi-layer perceptron (MLP), designed to embed BSM theories into a latent space such that phenomenologically similar models are clustered together. The encoder leverages physical observables like missing transverse energy (MET) and jet characteristics, aggregating these observables over multiple Monte Carlo-simulated events to form the input vector.

A contrastive loss function guides the embedding process based on whether pairs of event sets belong to the same or different models. This approach prioritizes separating dissimilar models and clustering similar ones, thus encoding the experimental indistinguishability into the resultant latent space.

Application and Results

The method is validated through three distinct scenarios:

1. MSSM Gluino Simplified Model: This scenario uses nine MSSM BSM models, each with different gluino and neutralino masses. The latent space effectively clustered the models according to the gluino-neutralino mass difference, indicating the dominant role of this parameter in determining the experimental signatures. All distinct model configurations were properly separated within the latent space with no overlapping regions for experiments with notably different mass gaps.
2. Dark Matter Simplified Models: Here, three dark matter production models are considered: vector mediator, pseudoscalar mediator, and squark mediator. Results showed that the latent space distinctly grouped vector and pseudoscalar models, while squark models formed separate clusters based on the mass difference between the squark and the dark matter candidate. Notably, the model clusters were ordered more sensitively according to the mass differences in the squark mediator model than in the vector and pseudoscalar models.
3. Dark Machines Anomaly Challenge Dataset: This scenario emphasizes practical complexity, incorporating several distinct BSM models along with Standard Model (SM) backgrounds. The latent space analysis successfully separated SM backgrounds from BSM signals and formed distinct clusters for different BSM processes based on MET and jet-related observables. Models such as the stop and vector mediator, with similar MET distributions, were grouped together but were still discernible within the latent space due to differences in jet momenta.

Implications and Future Directions

The proposed method demonstrates the utility of ML in abstractly representing complex theory spaces in a way that respects experimental observables. The embedding approach has multiple practical and theoretical implications:

• Benchmarking: The technique aids in selecting a minimal set of benchmark models that best cover the observable signature space, reducing redundancy in searches.
• Model Reinterpretation: Simplified reinterpretation becomes feasible by projecting experimental results into the latent space and comparing with pre-existing clusters.
• Anomaly Detection: The method offers new avenues for model-agnostic anomaly detection, identifying unexplored regions of the latent space that may signify potential signatures of new physics.
• Inverse Problems: The spatial organization of models in latent space can facilitate solving inverse problems, predicting possible BSM scenarios from observed data.

Future efforts should focus on embedding more complex features, incorporating cross-section information, and extending this latent space methodology to dynamically sample and reconstruct unpopulated physical spaces. This could provide a more comprehensive map of potential discovery regions, guiding both current analyses at the LHC and planning for future colliders.

In conclusion, the paper validates a machine learning-based strategy for organizing high-dimensional BSM spaces into intuitive, low-dimensional latent spaces. This approach not only enhances our ability to distinguish between phenomenologically similar models but also ensures better coverage and interpretation of the signature space, marking a noteworthy advancement in the exploration of new physics at high-energy colliders.