MERIT

MERIT: compare materials & understand their environmental impact. Our AI chatbot uses a knowledge graph and real-time data retrieval to give answers & visualizations. Material selection made easy!

Comment

Inspiration

MERIT stands for Materials Education, Research, and Informatics Tool. Inspired by the need for accurate and reliable information in materials science, MERIT aims to address the limitations of traditional Large Language Models (LLMs) that are prone to "hallucination" - generating convincing but false information.

What it does

MERIT provides on-demand solutions and insights about materials by integrating You.com AI search engine, high-fidelity computational materials database Materials Project, and Neo4j knowledge graph database.

  • Material Informatics: Users can retrieve more accurate materials information, inspect and and compare retrieved information through knowledge graph.
  • Environmental Impact: MERIT helps users understand the environmental footprint of various materials, supporting sustainability efforts.
  • Data-driven Predictions: The chatbot leverages a knowledge graph, You.com, research articles, and Materials Project data to provide accurate values for material properties like bulk modulus, bandgap, magnetic orderings, formation energies, and even synthesis procedures.
  • Visualizations: Complex material data is made accessible through intuitive knowledge graph visualizations, simplifying analysis and decision-making.

How we built it

MERIT is built using a multi-modal Retrieval-Augmented Generation (RAG) framework with hierarchical reasoning-and-acting (ReAct) agents. This means:

  • Knowledge Graph: A core component storing structured information about materials and their relationships.
  • Real-Time Data Retrieval: The system pulls information on-demand from external knowledge sources like You.com, research articles, and Materials Project API.
  • ReAct Agent: Agent within MERIT interacts with data sources, performs computations, and runs simulations using tools like atomistic workflows.
  • Hierarchical Planning: Multiple ReAct agents work together under the supervision of a main agent, allowing for complex reasoning and integration of diverse data modalities.

Challenges we ran into

  • Ensuring Data Accuracy: Integrating data from multiple sources requires careful validation and consistency checks to mitigate potential errors.
  • LLM Hallucination: Preventing the model from generating false information requires grounding it firmly in reliable data sources.
  • Handling Complex Queries: Understanding multi-part questions and performing multi-step reasoning with external tools is a challenging task.

Accomplishments that we're proud of

  • Successful Demo
  • Successful You.com and AWS Bedrock
  • Neo4j Knowledge Graph

What we learned

  • You.com and AWS Bedrock from amazing and informative seminars
  • UI/UX Design and Neo4j Knowledge Graph [converting Neo4j to Visnet]
  • Teamwork

What's next for MERIT

  • Expanding Knowledge Sources: Integrating additional databases and data sources to provide a more comprehensive view of materials.
  • Improving User Interface: Developing a more interactive and user-friendly interface to enhance the user experience.
  • Exploring New Applications: Applying MERIT to new research areas and tasks, such as material discovery, property prediction, and autonomous laboratory experiments.

Built With

Share this project:

Updates