Abstract
Particle physics theories, such as those which explain neutrino flavor mixing, arise from a vast landscape of model-building possibilities. A model's construction typically relies on the intuition of theorists. It also requires considerable effort to identify appropriate symmetry groups, assign field representations, and extract predictions for comparison with experimental data. We develop an Autonomous Model Builder (AMBer), a framework in which a reinforcement learning agent interacts with a streamlined physics software pipeline to search these spaces efficiently. AMBer selects symmetry groups, particle content, and group representation assignments to construct viable models while minimizing the number of free parameters introduced. We validate our approach in well-studied regions of theory space and extend the exploration to a novel, previously unexamined symmetry group. While demonstrated in the context of neutrino flavor theories, this approach of reinforcement learning with physics software feedback may be extended to other theoretical model-building problems in the future.
Introduction
In particle physics, the Standard Model remains incomplete, particularly concerning the origin of neutrino masses and their specific "flavor-mixing" patterns. Traditionally, theorists address this by constructing "flavor models" through the selection of mathematical symmetry groups and the assignment of particles to specific representations. However, this process faces significant hurdles: a discrete and categorical search space, high dimensionality with trillions of possible assignments, and high computational costs for evaluating each model. Historically, this has relied on human intuition, leaving vast regions of the theoretical landscape unexplored.
The AMBer Framework
To automate this "intuition-heavy" process, researchers have introduced AMBer (Autonomous Model Builder). AMBer utilizes Reinforcement Learning (RL), specifically the Proximal Policy Optimization (PPO) algorithm, to act as an autonomous theorist.
The framework consists of two primary components:
- The Agent: A neural network that takes actions such as adding particles, modifying symmetry representations, or adjusting vacuum expectation values (VEVs).
- The Environment: A custom scientific software pipeline. This includes PyDiscrete, a tool significantly faster than previous implementations, and Model2Mass, which automates the conversion of symmetry assignments into mass matrices.
The agent is guided by a reward function that prioritizes accuracy (low $\chi^2$ fit to experimental data from NuFIT v5.3) and elegance (minimal free parameters), while penalizing physically invalid models.
Key Insights and Results
AMBer successfully integrates symbolic theory construction into the RL loop, allowing the agent to "learn" group theory rules through trial and error. By balancing exploration and exploitation, the agent can navigate the high-dimensional theory space more effectively than traditional random scans.
Rediscovery and Efficiency
In the well-studied $A_4$ symmetry space, AMBer rediscovered model patterns previously identified by human theorists over decades. Furthermore, it demonstrated superior efficiency; in the $A_4 \times \mathbb{Z}_N$ space, AMBer identified 4,681 "good" models (defined as $\chi^2 \le 10$ and $\le 7$ parameters), compared to only 30 found via random sampling.
Beyond Human Intuition
The agent was also applied to the $T_{19}$ symmetry group, a complex space largely overlooked by human researchers. AMBer successfully identified multiple high-quality models in this novel territory, including one with only four free parameters that achieved a remarkably good fit ($\chi^2 \approx 10$). This highlights the potential for AI to uncover elegant theories that might elude human intuition.
Conclusion and Future Work
AMBer demonstrates that RL can effectively automate complex theoretical model-building in particle physics. While currently limited by simplified physics (e.g., excluding scalar potentials and RGE running) and a fixed scope (seesaw mechanisms), future iterations could incorporate Large Language Models for natural-language interaction and expand to other domains such as Dark Matter or Inflationary Cosmology.
Summary updated