[Paper Review] 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+3 more|arXiv (Cornell University)|Aug 21, 2024

Catalytic Processes in Materials ScienceMaterials Science3 citations

TL;DR

This paper introduces a multi-agent workflow that leverages chemistry foundation models and cross-modal retrieval via OpenCLIP to enable structure-focused, semantic information retrieval for small molecules, polymers, and reactions. By integrating latent representations from foundation models into retrieval-augmented generation, the system achieves improved accuracy and efficiency in catalyst and materials design tasks, demonstrating state-of-the-art performance in complex, natural language-driven queries.

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 autonomous 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 agentic systems on the retrieval of salient information for material design tasks. Within this context, alternative uses of predictive deep learning models, such as leveraging their latent representations to facilitate cross-modal retrieval augmented generation within agentic systems for task-specific materials design, has remained unexplored. Herein, we demonstrate that large, pre-trained chemistry foundation models can serve as a basis for enabling structure-focused, 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 multi-modal models such as OpenCLIP facilitate unprecedented queries and information retrieval across multiple characterization data domains. Finally, we demonstrate the integration of these models within multi-agent systems to facilitate structure and topological-based natural language queries and information retrieval for different research tasks.

Motivation & Objective

• To address the limitations in information retrieval within agentic systems for materials design, particularly in retrieving semantically relevant structural data.
• To explore the use of latent representations from pre-trained chemistry foundation models for cross-modal retrieval across molecular structures, reactions, and characterization data.
• To integrate these representations into multi-agent workflows that support natural language queries based on molecular topology and structure.
• To improve retrieval-augmented generation in materials discovery by enabling precise, semantic retrieval of relevant chemical data.
• To demonstrate the effectiveness of combining foundation models with multi-modal models like OpenCLIP for complex, structure-based queries in materials science.

Proposed method

• Utilizes large pre-trained chemistry foundation models to extract semantic representations of molecular structures, reactions, and polymeric materials.
• Employs cross-modal retrieval by aligning text queries with molecular structure embeddings using OpenCLIP, enabling semantic search across diverse data modalities.
• Designs a multi-agent system where specialized agents handle query interpretation, retrieval, and generation, leveraging the foundation model's latent space for accuracy.
• Integrates retrieval-augmented generation (RAG) with structure-focused embeddings to improve response relevance in materials design tasks.
• Trains and fine-tunes the retrieval pipeline using multi-modal contrastive learning to align text and structural representations across different data domains.
• Applies the framework to real-world catalyst and materials design tasks, supporting natural language queries about topology and chemical structure.

Experimental results

Research questions

• RQ1 Can chemistry foundation models improve semantic, structure-focused retrieval in multi-agent systems for materials design?
• RQ2 How effective is the integration of foundation model embeddings with multi-modal models like OpenCLIP for cross-modal retrieval in materials science?
• RQ3 To what extent does structure-aware retrieval enhance the quality of generative outputs in catalyst and materials design workflows?
• RQ4 Can multi-agent systems leverage latent representations from foundation models to handle complex, natural language queries about molecular topology?
• RQ5 What performance gains are achieved in retrieval accuracy and generation relevance when using structure-focused RAG with foundation models?

Key findings

• The system achieves state-of-the-art performance in structure-focused retrieval for small molecules, polymers, and reactions using semantic embeddings from chemistry foundation models.
• Integration with OpenCLIP enables unprecedented cross-modal retrieval across diverse characterization data domains, including spectroscopic and structural data.
• The multi-agent workflow significantly improves retrieval accuracy and relevance in response to complex, natural language queries about molecular topology and structure.
• Retrieval-augmented generation using foundation model embeddings leads to more chemically plausible and targeted materials design suggestions.
• The approach demonstrates robust generalization across multiple materials classes, including complex polymeric systems and catalytic materials.
• Quantitative evaluation shows measurable improvements in retrieval precision and generation quality compared to baseline RAG systems without structure-aware embeddings.