Bayesian active search on parameter space: A 95 GeV spin-0 resonance in the ($B-L$)SSM

Mauricio A. Diaz, Giorgio Cerro, Srinandan Dasmahapatra, Stefano MorettiSciPost 2025 · DOI

TL;DR

This paper introduces b-CASTOR, a novel machine learning approach for efficiently exploring complex theoretical models in particle physics. It significantly outperforms traditional methods like MCMC in finding parameter regions that explain experimental anomalies, specifically a potential 95 GeV Higgs boson signal within the (B-L)SSM framework.

What Problem Does This Paper Address?

Particle physics theories beyond the Standard Model (BSM) often have vast parameter spaces. To test these theories against experimental data, researchers need to scan these spaces to find parameter configurations that accurately predict observed phenomena. This is computationally very expensive, especially when multiple experimental observables (like signal strengths for different decay channels) need to be simultaneously satisfied. Traditional methods like Markov-Chain Monte Carlo (MCMC) struggle with this "curse of dimensionality" and can be inefficient, requiring millions of evaluations. This paper tackles the challenge of efficiently finding "satisfactory" regions in these high-dimensional, computationally expensive parameter spaces, particularly when dealing with multiple, potentially sparse, and disconnected regions of interest.

Methods

The core contribution is the b-CASTOR algorithm, a batched Constrained Active Search method. It combines several key ideas:

  • Probabilistic Surrogate Models: Instead of directly evaluating the complex (B-L)SSM model for every parameter point, b-CASTOR uses Gaussian Processes (GPs) to build "surrogate models." These models learn to approximate the model's predictions (observables) based on a limited number of evaluations, quantifying uncertainty in their predictions.
  • Expected Coverage Improvement (ECI) Policy: This policy guides the search by aiming to maximize the volume of the "satisfactory" region (where experimental constraints are met) that the surrogate model covers. It balances exploring unknown areas with exploiting areas likely to contain solutions.
  • Tree-structured Parzen Estimator (TPE): TPE is used to optimize the ECI policy efficiently, especially for selecting multiple points at once.
  • Rank-based Sampling & Batch Evaluation: Instead of evaluating points one by one, b-CASTOR samples points in batches. A rank-based strategy prioritizes points with higher ECI values, allowing for more efficient exploration and exploitation of the parameter space.
  • Decaying Radius: The search radius around promising points decreases over iterations, allowing for initial broad exploration and then fine-grained resolution filling of the satisfactory region.

Datasets

The paper uses two main "datasets" for evaluation:

  1. A 2D Double-Objective Test Function: A synthetic function (fBH) designed to mimic the complexity of sparse and disconnected satisfactory regions in multi-objective optimization. This serves as a benchmark for testing the algorithm's core mechanics.
  2. The (B-L)SSM Model: This is the primary focus. The "dataset" here refers to the output of a complex High Energy Physics (HEP) software stack (SPheno, MadGraph, HiggsBounds, HiggsSignals) that simulates the (B-L)SSM model's predictions for various parameter configurations. This stack is computationally expensive, taking minutes per evaluation.

Key Insights and Ideas

  • Active Search for Efficiency: The core idea is that by intelligently selecting where to evaluate the expensive model, we can find solutions much faster than random or brute-force methods.
  • Multi-Objective Optimization is Key: Real-world physics problems often involve satisfying multiple experimental constraints simultaneously. b-CASTOR is designed to handle this complexity, unlike single-objective methods.
  • Surrogate Models as Smart Guides: GPs provide not just predictions but also uncertainty estimates, which are crucial for guiding the search effectively.
  • Batching for Parallelism: Evaluating multiple points simultaneously significantly speeds up the overall search process, especially on modern multi-core processors.
  • The (B-L)SSM as a Challenging Testbed: The chosen theoretical model is intentionally complex, with a large parameter space and computationally demanding predictions, making it an ideal, albeit difficult, test case.

Key Results

  • Superior Sample Efficiency: In the 2D test function case, b-CASTOR found an average of 2090 satisfactory points out of 2200 calls (94.57% efficiency), while MCMC-MH found only 338 (15.29%).
  • Effective Characterization of Satisfactory Regions: For the (B-L)SSM, b-CASTOR identified 1636 satisfactory configurations (50% efficiency) within 3240 calls, effectively characterizing the region explaining the 95 GeV diphoton anomaly. MCMC-MH found only 25 satisfactory points in the same number of calls.
  • Inability to Simultaneously Explain All Anomalies: When attempting to explain the 95 GeV anomaly across three channels (γγ, ττ, and bb), b-CASTOR could not find any parameter configurations that satisfied all three experimental constraints simultaneously. However, it did identify nine points that satisfied two of the three.
  • Robustness: The algorithm showed similar efficiency in another model (2HDM), suggesting broad applicability.

Limitations and Future Work

  • Computational Cost: While significantly more efficient than MCMC, evaluating the full HEP software stack remains computationally intensive.
  • Inability to Explain All Three Anomalies: The paper highlights that the (B-L)SSM, within the explored parameter space, cannot simultaneously accommodate the γγ, ττ, and bb anomalies. This suggests either the model needs modification, the parameter space needs further exploration, or these anomalies might not originate from a single source within this model.
  • Future Work: The authors suggest exploring Bayesian Neural Networks (BNNs) as surrogate models for potentially faster inference and better scaling to larger datasets. They also mention the possibility of incorporating more sophisticated exploration strategies and comparing b-CASTOR against other state-of-the-art ML-based scanning algorithms.

Summary updated

Abstract

In the attempt to explain possible data anomalies from collider experiments in terms of New Physics (NP) models, computationally expensive scans over their parameter spaces are typically required in order to match theoretical predictions to experimental observations. Under the assumption that anomalies seen at a mass of about 95 GeV by the Large Electron-Positron (LEP) and Large Hadron Collider (LHC) experiments correspond to a NP signal, which we attempt to interpret as a spin-0 resonance in the (B-L) <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"> <mml:mrow> <mml:mo stretchy="false" form="prefix">(</mml:mo> <mml:mi>B</mml:mi> <mml:mo>−</mml:mo> <mml:mi>L</mml:mi> <mml:mo stretchy="false" form="postfix">)</mml:mo> </mml:mrow> </mml:math> Supersymmetric Standard Model ( (B-L) <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"> <mml:mrow> <mml:mo stretchy="false" form="prefix">(</mml:mo> <mml:mi>B</mml:mi> <mml:mo>−</mml:mo> <mml:mi>L</mml:mi> <mml:mo stretchy="false" form="postfix">)</mml:mo> </mml:mrow> </mml:math> SSM), chosen as an illustrative example, we introduce a novel Machine Learning (ML) approach based on a multi-objective active search method, called b-CASTOR, able to achieve high sample efficiency and diversity, due to the use of probabilistic surrogate models and a volume based search policy, outperforming competing algorithms, such as those based on Markov-Chain Monte Carlo (MCMC) methods.
BibTeX
@inproceedings{diaz2025bayesian,
  author = {Mauricio A. Diaz and Giorgio Cerro and Srinandan Dasmahapatra and Stefano Moretti},
  title = {{Bayesian active search on parameter space: A 95 GeV spin-0 resonance in the ($B-L$)SSM}},
  booktitle = {SciPost},
  year = {2025},
  doi = {10.21468/scipostphyscore.8.4.068},
}
Figure