Abstract
This paper presents hep-aid, a modular Python library designed for sample-efficient, multi-objective parameter scans in Beyond the Standard Model (BSM) phenomenology. By integrating Machine Learning (ML) surrogate models with Active Search (AS) algorithms, hep-aid addresses the computational bottleneck of high-dimensional parameter space exploration. The library includes implementations of Constraint Active Search (CAS), b-CASTOR, and MLScan, alongside a standard MCMC-MH baseline. Demonstrations on synthetic benchmarks and real-world BSM scenarios show significant improvements in search efficiency compared to traditional methods.
Introduction
In the search for "new physics," researchers explore Beyond the Standard Model (BSM) theories to address unresolved questions like dark matter and the hierarchy problem. These theories introduce numerous unknown parameters, such as masses and coupling constants. Validating these models requires "Parameter Scans" (PS), where points in the parameter space are evaluated against experimental data from facilities like the Large Hadron Collider (LHC).
However, these evaluations involve complex simulation "stacks" that are computationally expensive. Traditional brute-force methods, such as random or grid sampling, are inefficient in high-dimensional spaces (often 8D or more), as they spend excessive resources on regions that do not satisfy physical constraints. hep-aid provides a framework to minimize the number of full simulations required by identifying "satisfactory" regions with high sample efficiency.
Library Architecture
The hep-aid library is built on a modular design, primarily divided into two functional components: the hep module and the search module.
The hep module serves as an interface to external High Energy Physics (HEP) software. It manages the execution of the "HEP-stack," which typically includes tools like SPheno for spectrum calculation, HiggsBounds/HiggsSignals for experimental constraints, and MadGraph for event generation. It abstracts the handling of SLHA (SUSY Les Houches Accord) files, allowing users to interact with these tools via standard Python dictionaries.
The search module implements the search logic. It utilizes Active Search (AS) strategies where a "surrogate model"—a fast ML approximation such as a Gaussian Process (GP) or Neural Network (NN)—is trained on the results of previous simulations. This surrogate predicts the likelihood of new points satisfying the objective constraints, guiding the search toward the most promising regions.
Search Algorithms
hep-aid integrates several advanced ML-based search methods:
- b-CASTOR: A batched Active Search method that leverages GPs and Tree-structured Parzen Estimators. It is designed for parallel execution in high-performance computing environments.
- MLScan: Utilizes a Multi-Layer Perceptron (MLP) surrogate and rejection sampling, excelling at exploring disconnected valid regions.
- CAS (Constraint Active Search): A sequential version of the active search logic focused on satisfying specific constraints.
- MCMC-MH: A standard Metropolis-Hastings algorithm provided as a baseline for performance comparison.
Benchmarking and Case Studies
The library's performance was evaluated using both synthetic functions and a complex BSM case study. Synthetic benchmarks, such as the "Egg Box" function, allow for the visualization of how algorithms navigate multi-modal landscapes with multiple "islands" of validity.
For a real-world application, the authors performed an 8-dimensional scan of the $(B-L)$ Supersymmetric Standard Model $((B-L)SSM)$. The objective was to find regions consistent with the observed 125 GeV Higgs mass while also explaining potential anomalies, such as the 95 GeV di-photon excess.
Key Insights and Results
The primary advantage of hep-aid is the use of surrogate models to steer the search. By learning the general shape of the constraint landscape, the library avoids wasting time on "dead" regions of the parameter space.
Key results from the study include:
- High Sample Efficiency: In the 8D BSM scan, the
b-CASTORalgorithm achieved an efficiency of nearly 95%, meaning the vast majority of points selected for full simulation met the "satisfactory" criteria. - Robustness: ML-based approaches significantly outperformed MCMC baselines, which often struggled with poor initialization and failed to discover all valid regions.
- Exploration vs. Exploitation: The library allows users to choose algorithms based on their specific needs.
b-CASTORis highly efficient at mapping out known regions (exploitation), whileMLScanis more robust at discovering new, disconnected regions (exploration).
Conclusion and Future Work
hep-aid offers a versatile and efficient toolset for BSM phenomenology, significantly reducing the computational cost of parameter scans. While currently integrated with the SARAH/SPheno ecosystem, future development aims to expand compatibility with a broader range of HEP software. Additionally, the authors envision hybrid approaches that combine the exploratory power of MLScan with the exploitative efficiency of b-CASTOR to further optimize the search for new physics.
Summary updated