HEP Analysis Multi-Agent System

A comprehensive multi-agent framework built on smolagents for High Energy Physics (HEP) analysis, designed for particle physics research including new particle searches and hypothesis verification.

Overview

This system implements a hierarchical multi-agent architecture where specialized agents handle different aspects of a physics analysis:

                    ┌─────────────────────────────┐
                    │    HEP Analysis Manager     │
                    │       (Orchestrator)        │
                    └──────────────┬──────────────┘
                                   │
        ┌──────────────────────────┼──────────────────────────┐
        │                          │                          │
        ▼                          ▼                          ▼
┌───────────────┐        ┌─────────────────┐        ┌────────────────┐
│ Infrastructure│        │ Physics Analysis │        │ Interpretation │
│    Agents     │        │     Agents       │        │    Agents      │
└───────────────┘        └─────────────────┘        └────────────────┘
        │                          │                          │
   ┌────┴────┐            ┌────────┴────────┐          ┌──────┴──────┐
   │         │            │                 │          │             │
   ▼         ▼            ▼                 ▼          ▼             ▼
┌─────┐  ┌─────┐    ┌─────────┐      ┌──────────┐  ┌────────┐  ┌────────┐
│ Web │  │Data │    │ Event   │      │ Fitting  │  │Uncert. │  │ Paper  │
│Srch │  │Retr │    │Selection│ ...  │& Stats   │  │Quant.  │  │Writing │
└─────┘  └─────┘    └─────────┘      └──────────┘  └────────┘  └────────┘

Agent Types

Infrastructure Agents

1. Web Research Agent - Searches arXiv, INSPIRE-HEP for papers and measurements
2. Data Retrieval Agent - Downloads data from Zenodo, CERN Open Data
3. Code Search Agent - Finds analysis code on GitHub/GitLab

Physics Analysis Agents

4. Event Selection Agent - Applies kinematic cuts and selection criteria
5. Data Processing Agent - Handles calibrations and derived variables
6. Reconstruction Agent - Reconstructs composite particles (Z, W, H)
7. Fitting Agent - Fits histograms with signal/background models
8. Statistical Analysis Agent - Hypothesis tests, significance, limits
9. Plotting Agent - Creates publication-quality plots (CMS/ATLAS style)
10. Uncertainty Agent - Calculates systematic/statistical uncertainties
11. Interpretation Agent - Interprets results and writes paper sections

Installation

# Clone the repository
cd /path/to/agentification_of_analysis

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or
.\venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

Configuration

Set up your API keys in a .env file:

# Choose one LLM provider
ANTHROPIC_API_KEY=your_anthropic_key
OPENAI_API_KEY=your_openai_key
HUGGINGFACE_TOKEN=your_hf_token

# Optional: for data access
ZENODO_TOKEN=your_zenodo_token
GITHUB_TOKEN=your_github_token

Quick Start

1. Interactive Mode

python main.py --interactive

2. Run Example Analysis

python main.py --example

3. Run Specific Task

python main.py --task "Search for Z' boson decaying to dimuon pairs"

4. Python API

from smolagents import LiteLLMModel
from agents import HEPAnalysisOrchestrator

# Initialize model
model = LiteLLMModel(model_id="anthropic/claude-sonnet-4-20250514")

# Create orchestrator
orchestrator = HEPAnalysisOrchestrator(model)

# Run analysis
result = orchestrator.run("Measure Z boson cross section in dimuon channel")

# Or run structured workflow
config = {
    "hypothesis": "Z boson production",
    "channel": "dimuon",
    "selection": {"pt": 25, "eta": 2.4}
}
results = orchestrator.run_workflow(config)

Analysis Workflow

The system follows a typical HEP analysis workflow:

1. Literature Review - Search for previous measurements and theory predictions
2. Data Retrieval - Download collision data and MC samples
3. Event Selection - Apply cuts to select signal events
4. Reconstruction - Build physics objects (leptons, jets, MET)
5. Fitting - Extract signal yield using fit models
6. Uncertainties - Evaluate systematic and statistical errors
7. Statistical Analysis - Calculate significance or limits
8. Interpretation - Compare with predictions, write results

Tool Reference

Physics Tools

Tool	Description
EventSelectionTool	Apply kinematic cuts and selection
ReconstructionTool	Reconstruct composite particles
HistogramFitterTool	Fit distributions with various models
StatisticalFitTool	Hypothesis tests and limit setting
SystematicUncertaintyTool	Calculate systematic uncertainties
PublicationPlotTool	Create CMS/ATLAS style plots

Infrastructure Tools

Tool	Description
ArxivSearchTool	Search arXiv preprints
InspireHEPSearchTool	Search INSPIRE-HEP database
ZenodoDataTool	Download data from Zenodo
CERNOpenDataTool	Access CERN Open Data Portal
ROOTFileLoaderTool	Load and inspect ROOT files
GitHubSearchTool	Search for analysis code

Examples

Z Boson Cross Section Measurement

from agents import HEPAnalysisOrchestrator

orchestrator = HEPAnalysisOrchestrator(model)

result = orchestrator.run("""
    Measure the Z boson production cross section in the dimuon channel:
    1. Search for previous CMS/ATLAS measurements
    2. Apply standard dimuon selection (pT > 25 GeV, |η| < 2.4)
    3. Fit the invariant mass spectrum
    4. Calculate cross section with uncertainties
    5. Compare with NNLO prediction
""")

New Particle Search

result = orchestrator.run("""
    Search for a heavy Z' boson decaying to muon pairs:
    1. Define signal region for high-mass dimuon events
    2. Estimate backgrounds from sidebands
    3. Perform bump hunt in mass spectrum
    4. Set upper limits if no excess found
""")

Project Structure

agentification_of_analysis/
├── agents/
│   ├── __init__.py
│   ├── infrastructure_agents.py
│   ├── physics_agents.py
│   └── orchestrator.py
├── tools/
│   ├── __init__.py
│   ├── infrastructure.py
│   ├── data_retrieval.py
│   ├── code_tools.py
│   ├── physics_tools.py
│   ├── fitting_tools.py
│   ├── plotting_tools.py
│   ├── uncertainty_tools.py
│   └── interpretation_tools.py
├── examples/
│   ├── __init__.py
│   └── z_boson_analysis.py
├── config.py
├── main.py
├── requirements.txt
└── README.md

Supported Models

• Anthropic Claude (recommended): claude-sonnet-4-20250514, claude-opus-4-20250514
• OpenAI: gpt-4, gpt-4-turbo
• HuggingFace: Qwen/Qwen2.5-72B-Instruct, meta-llama/Llama-3.3-70B-Instruct

References

• smolagents Documentation
• HuggingFace Agents Course
• CERN Open Data
• INSPIRE-HEP

License

MIT License