DarkAgents: Multi-Agent System Empowers Astroparticle Physics Research

DarkAgents organically combines the reasoning capabilities of large language models (LLMs) with deterministic human-written code to build an automated research pipeline. It was first applied to the study of cosmological first-order phase transitions, capable of outputting optimal fitting parameters, experimental constraints, and hypothesis audit reports, providing a new paradigm for astroparticle physics research.

Astroparticle physics studies the most fundamental components and interactions of the universe, covering cutting-edge topics such as dark matter, dark energy, gravitational waves, and early universe phase transitions. Research in this field is highly complex: theoretical models involve a large number of free parameters, computational workflows require chaining together multiple professional software tools, and experimental data comes from different observation methods and needs to meet various physical constraints. The traditional research model relies on manual operations based on individual experience, which easily introduces human errors. Moreover, incomplete records of assumptions and prior choices lead to difficulty in reproducing and auditing results, a problem that is particularly prominent in cross-disciplinary collaborations.

DarkAgents is a multi-agent system specifically designed for astroparticle physics, with its core innovation being the integration of large language model (LLM) reasoning and deterministic human-written code. The system adopts a modular architecture, supporting multiple LLM backends such as Mistral, Anthropic, OpenAI API, and local models. The workflow consists of three stages: Model Construction (agents generate theoretical model code based on physical hypotheses), Computational Pipeline (automatically orchestrates steps like differential equation solving, numerical integration, and Monte Carlo simulation), and Constraint Validation (compares computational results with existing experimental data to identify allowed parameter regions). LLMs are responsible for understanding physical intent, generating code frameworks, and handling exceptions; core computations are executed using tested human-written code to ensure scientific rigor.

The first application case is the study of cosmological first-order phase transitions (closely related to dark matter generation, the origin of baryon asymmetry, and gravitational wave production). The research starts from a classical scale-invariant particle physics model and fits NANOGrav nanohertz gravitational wave spectrum data (NANOGrav is the North American Nanohertz Observatory for Gravitational Waves, which detects low-frequency gravitational waves through pulsar arrival time measurements). The outputs include: optimal fitting values of model parameters, existing experimental observation constraints, and a hypothesis and prior audit report (detailing the assumptions in the analysis process and their impact on results). Tests revealed inconsistencies in some fitting results in the literature, and new fitting results were generated based on dissipative fluid gravitational wave templates. The project code is open-sourced on GitHub (https://github.com/PhysicsZandi/DarkAgents).

Test runs revealed inconsistencies in some fitting results in the literature (caused by differences in computational assumptions, numerical methods, or data processing). DarkAgents' automated audit function helps identify and locate these issues; new fitting results were generated based on the dissipative fluid mechanism; the open-source code provides a foundation for other researchers to reproduce results, extend functions, and apply the system to new topics.

DarkAgents represents an important direction for AI agents in basic science applications, deeply customized to meet specific domain needs and balancing automation and reliability. It can be extended to high-energy physics, cosmology, nuclear physics, and other fields facing similar challenges. For the astroparticle physics community, it is a new research paradigm that emphasizes transparency, reproducibility, and systematic hypothesis management. With the improvement of LLM capabilities and the accumulation of domain knowledge, AI agents are expected to play a key role in more basic science fields.