AgenticSciML

Collaborative Multi-Agent Systems for Emergent Discovery in Scientific Machine Learning

Over 10 specialized AI agents that propose, critique, and evolve SciML solutions through structured reasoning, retrieval-augmented method memory, and ensemble-guided evolutionary search.

Qile Jiang · George Em Karniadakis

Paper: https://www.nature.com/articles/s44387-026-00102-5

• 10+ specialized agents collaborate via Multi-Agent Debate — proposing, critiquing, and refining candidate solutions.
• Retrieval-augmented method memory (RAG) grounds the agents in a curated knowledge base of SciML literature and techniques.
• Ensemble-guided evolutionary tree search mutates and selects solutions, balancing exploration and exploitation.
• Up to 4 orders of magnitude error reduction over single-agent baselines.
• Emergent, novel methods — adaptive mixture-of-expert architectures, decomposition-based PINNs, and physics-informed operator models that are not explicitly present in the knowledge base.

1 · Python packages.

pip install python-dotenv pydantic langgraph
pip install langchain-core langchain-anthropic langchain-openai langchain-google-genai
pip install torch numpy scipy matplotlib

2 · API keys. Create a .env file inside src/ with the providers you plan to use:

ANTHROPIC_API_KEY=your_anthropic_key    # Claude models
OPENAI_API_KEY=your_openai_key          # GPT models
GOOGLE_API_KEY=your_google_key          # Gemini models
XAI_API_KEY=your_xai_key                # Grok models

3 · Problem setup. Describe your problem in src/USER_INPUT/:

A ready-to-run function-approximation example ships in src/USER_INPUT/.

We recommend the three-step, human-in-the-loop pipeline:

cd src

# Step 1 — Generate & approve the testing contract (evaluate.py + guidelines.md in TESTING/)
python main.py --mode contract-only

# Step 2 — Synthesize, validate, and train the root solution (SOLUTION_AND_OUTPUTS/solution_0/)
python main.py --mode root-only

# Step 3 — Evolve solutions via mutation, selection, and evolution
python main.py --mode evolve-only

Prefer to run everything end-to-end (no manual approval — use with care)?

python main.py --mode full

Multi-GPU. Tree expansion is parallelized across GPUs:

python main.py --mode full --gpu_ids 0 1 2 3   # use 4 specific GPUs
python main.py --mode full                      # auto-detect GPUs (falls back to CPU)

💡 Best practice: generate and carefully review the contract first (locally is fine), then run root-only and evolve-only on a cluster — these phases are compute-intensive and benefit from GPUs. An example SLURM script is provided in src/main_run.sh.

AgenticSciML/
├── framework.jpg          # System overview figure
└── src/
    ├── main.py            # Pipeline entry point (contract / root / evolve / full)
    ├── agents.py          # Agent definitions, roles, and model routing
    ├── create_contract.py # Phase 2 — testing-contract generation
    ├── create_root.py     # Phase 3 — root-solution synthesis
    ├── propose_critic.py  # Multi-agent debate (proposal–critic loop)
    ├── select_mutations.py# Evolutionary mutation selection
    ├── retrieve_*.py      # Retrieval-augmented method memory (RAG)
    ├── analyze.py         # Solution-tree analysis
    ├── telemetry.py       # Run telemetry
    ├── KB/                # Curated SciML method knowledge base
    ├── USER_INPUT/        # Your problem spec + example dataset
    └── main_run.sh        # Example SLURM launch script

This code was developed and validated against the LLM generation available at the time of the paper (circa 2025). The agent prompts and model-interaction code have not been updated since.

If you run AgenticSciML against newer models, expect to do some retuning:

• Prompt drift — newer models may respond differently to the existing prompts; you may need to re-tune wording, formatting, or system instructions to recover the reported behavior.
• Sampling parameters — some newer APIs restrict or ignore parameters such as temperature (e.g., certain reasoning models). The per-agent temperatures in src/agents.py may need adjustment or removal.
• API changes — model IDs, client signatures, and provider SDKs evolve; parts of the model-routing code may need updating to match the latest LLM APIs.

We welcome PRs that modernize the prompts and provider integrations. 🙌

If you use this code or build on this work, please cite:

@article{jiang2026agenticsciml,
  title={Agenticsciml: Collaborative multi-agent systems for emergent discovery in scientific machine learning},
  author={Jiang, Qile and Karniadakis, George},
  journal={npj Artificial Intelligence},
  year={2026},
  publisher={Nature Publishing Group UK London}
}

Released under the MIT License — see LICENSE for details.