Towards Beyond Standard Model Model-Building with Reinforcement Learning on Graphs

TL;DR

This paper introduces a reinforcement learning framework using graph representations to automate the exploration and construction of beyond Standard Model theories, successfully identifying models that address current physics anomalies.

Contribution

The paper presents a novel graph-based reinforcement learning approach for automated model-building in particle physics, capable of exploring complex theory spaces without predefined particle counts.

Findings

Successfully identified models explaining muon g-2 anomaly

Discovered both known and new viable theories

Demonstrated effective exploration of continuous and discrete parameters

Abstract

We provide a framework for exploring physics beyond the Standard Model with reinforcement learning using graph representations of new physics theories. The graph structure allows for model-building without a priori specifying definite numbers of new particles. As a case study, we apply our method to a simple class of theories involving vectorlike leptons and a dark U(1) inspired by the portal matter paradigm. Using modern policy gradient methods, the agent successfully explores a model space consisting of both continuous and discrete parameters and identifies consistent theories. The minimal models found include both known and previously unstudied examples that can accommodate the muon anomalous magnetic moment and satisfy precision electroweak and flavor constraints. The method represents a step forward in enabling an automated model-building process for physics beyond the Standard Model.