SciAgents automates scientific discovery by employing multi-agent systems to reason over large-scale ontological knowledge graphs. It targets researchers seeking to uncover interdisciplinary relationships and generate novel hypotheses in fields like bio-inspired materials, offering a scalable and precise approach that surpasses traditional methods.

How It Works

The system integrates three core components: extensive ontological knowledge graphs for organizing scientific concepts, a suite of LLMs and data retrieval tools, and multi-agent systems with in-situ learning. Two frameworks are provided: a pre-programmed sequence of agent interactions for reliability and a dynamic, adaptive framework. Agents like an Ontologist, Scientist, Critic, Planner, and Assistant collaborate to generate, refine, and critique research hypotheses, leveraging graph paths for context.

Quick Start & Requirements

• Install via pip: pip install git+https://github.com/lamm-mit/GraphReasoning or pip install -e git+https://github.com/lamm-mit/GraphReasoning.git#egg=GraphReasoning.
• Requires OpenAI and Semantic Scholar API keys.
• May require wkhtmltopdf (sudo apt-get install wkhtmltopdf).
• Graph and embedding files are downloaded from Hugging Face Hub.
• See official notebooks: SciAgents_ScienceDiscovery_GraphReasoning_non-automated.ipynb and SciAgents_ScienceDiscovery_GraphReasoning_automated.ipynb.

Highlighted Details

• Implements both automated and non-automated multi-agent frameworks.
• Leverages AG2 (formerly AutoGen) for the automated multi-agent model.
• Generates detailed research hypotheses, including mechanisms, design principles, and novelty assessments.
• Demonstrates discovery of hidden interdisciplinary relationships in bio-inspired materials.

Maintenance & Community

• Developed by researchers at MIT (A. Ghafarollahi, M.J. Buehler).
• Code is available on GitHub.
• Related project for audio generation: lamm-mit/PDF2Audio.

Licensing & Compatibility

• The repository itself does not explicitly state a license. The underlying AG2 framework is Apache 2.0.
• API key requirements (OpenAI, Semantic Scholar) may have associated costs and usage terms.

Limitations & Caveats

The project requires API access to OpenAI and Semantic Scholar, which may incur costs and have usage restrictions. The effectiveness of the discovery process is dependent on the quality and comprehensiveness of the underlying knowledge graph and the LLMs used.