Machine learning methods for exploring beyond standard model parameters

Abstract

Physics aims to describe natural phenomena through the construction of theoretical models that capture essential system behaviours, with predictions tested against observations. In particle physics, the Standard Model (SM) has been a monumental success but remains incomplete, leaving unresolved challenges such as explaining neutrino masses, dark matter, the hierarchy problem, and gravity. Moreover, recent experimental anomalies, including results from scalar searches, hint at new physics and motivate the exploration of Beyond the SM (BSM) scenarios.

However, performing phenomenological studies in BSM models poses two major challenges. First, the number of possible models is immense. Second, within a single model, the parameter space is characterised by high dimensionality, sparsity of feasible configurations, and the computational cost of numerical evaluations, necessitating advanced parameter scan algorithms.

This thesis introduces a new formulation for parameter scan algorithms based on an active search methodology. This approach leverages Machine Learning (ML) modelling and sequential decision-making techniques, borrowing concepts from Bayesian Optimisation. A new, sample-efficient parameter scan algorithm, called b-CASTOR, is proposed. Additionally, a Python library named \hepaid\ is presented, designed for the easy use, integration, and development of parameter scan algorithms in phenomenological studies. Finally, a Reinforcement Learning formulation for parameter space scans is reviewed. While this approach yielded negative results, the insights and limitations derived from the project are discussed.

This thesis aims to advance the development of ML-based parameter scan algorithms, addressing computational challenges and laying the foundations for a systematic exploration of BSM models.

More information

Identifiers

ORCID for Stefano Moretti: orcid.org/0000-0002-8601-7246

ORCID for Srinandan Dasmahapatra: orcid.org/0000-0002-9757-5315

Catalogue record

Export record