Leveraging Chemistry Foundation Models to Facilitate Structure Focused Retrieval Augmented Generation in Multi-Agent Workflows for Catalyst and Materials Design

Nathaniel H. Park,Tiffany J. Callahan,James L. Hedrick,Tim Erdmann,Sara Capponi

2024-08-22

Abstract:Molecular property prediction and generative design via deep learning models has been the subject of intense research given its potential to accelerate development of new, high-performance materials. More recently, these workflows have been significantly augmented with the advent of large language models (LLMs) and systems of LLM-driven agents capable of utilizing pre-trained models to make predictions in the context of more complex research tasks. While effective, there is still room for substantial improvement within the agentic systems on the retrieval of salient information for material design tasks. Moreover, alternative uses of predictive deep learning models, such as leveraging their latent representations to facilitate cross-modal retrieval augmented generation within agentic systems to enable task-specific materials design, has remained unexplored. Herein, we demonstrate that large, pre-trained chemistry foundation models can serve as a basis for enabling semantic chemistry information retrieval for both small-molecules, complex polymeric materials, and reactions. Additionally, we show the use of chemistry foundation models in conjunction with image models such as OpenCLIP facilitate unprecedented queries and information retrieval across multiple characterization data domains. Finally, we demonstrate the integration of these systems within multi-agent systems to facilitate structure and topological-based natural language queries and information retrieval for complex research tasks.

Artificial Intelligence

What problem does this paper attempt to address?

The paper aims to address the following issues:

1. **Structural Similarity Retrieval**: Existing multi-agent systems still lack effective information retrieval capabilities in material design tasks, especially in terms of structural similarity-based information retrieval. The paper demonstrates that large pre-trained chemical foundation models (such as MoLFormer) can be used for semantic chemical information retrieval of small molecules, complex polymers, and reactions.

2. **Cross-Modal Retrieval Augmented Generation**: Exploring how to utilize the latent representations of chemical foundation models to promote Retrieval-Augmented Generation (RAG) for multimodal retrieval, thereby achieving material design for specific tasks. For example, combining image models (such as OpenCLIP) for querying and information retrieval across representational data domains.

3. **Multi-Agent System Integration**: Integrating these systems into multi-agent workflows to support natural language queries and information retrieval based on structure and topology, thereby enhancing the capabilities for complex research tasks.

Through these methods, the paper aims to significantly enhance the functionality of multi-agent systems, particularly in material design and development tasks. Specifically, the authors validate the effectiveness of MoLFormer embeddings in different similarity search strategies and further extend its application to polymers and reactions, as well as achieving cross-modal queries by combining with image models. Ultimately, by constructing a hierarchical multi-agent self-reflective RAG system, effective handling of user queries and comprehensive report generation is achieved.