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

Leveraging Chemistry Foundation Models for Structure-Focused RAG in Multi-Agent Workflows

Abstract

The integration of Large Language Models (LLMs) into chemical research has led to the development of autonomous agents capable of complex task execution. However, traditional Retrieval Augmented Generation (RAG) systems often rely on natural language embeddings, which fail to capture the nuanced structural and topological similarities critical to materials science. This paper presents a novel approach by leveraging chemistry foundation models, specifically MoLFormer, to enable structure-focused, semantic information retrieval. We demonstrate that these structural embeddings facilitate "chemical reasoning" through vector arithmetic and multi-modal integration with characterization data like NMR spectra. Integrated within a hierarchical multi-agent workflow, this system significantly enhances the retrieval of salient information for catalyst and polymer design.

Introduction

In the rapidly evolving field of AI-driven materials discovery, LLMs serve as central orchestrators for autonomous research agents. While effective at processing textual data, these agents encounter a significant bottleneck when tasked with retrieving chemically relevant information. Standard RAG pipelines utilize text-based embeddings that prioritize lexical similarity over chemical structure. Traditional cheminformatics alternatives, such as molecular fingerprints, are often too rigid to capture the "semantic" similarity that human chemists intuitively recognize. This work addresses these limitations by utilizing latent representations from large-scale chemistry foundation models to bridge the gap between structural chemistry and agentic workflows.

Methods

The core of our approach is the implementation of a structure-focused RAG system. We utilized MoLFormer, a transformer-based model pre-trained on millions of SMILES strings, to generate high-dimensional structural embeddings. These embeddings allow for "chemical vector arithmetic," enabling operations such as functional group swapping via simple latent space manipulation (e.g., Catalyst A - Group X + Group Y).

To handle multi-modal characterization data, we paired MoLFormer with OpenCLIP, a computer vision model. By embedding images of NMR spectra and cross-referencing them with structural embeddings in a Milvus vector database, we enabled cross-modal queries. The entire system is orchestrated by a hierarchical multi-agent framework built using LangGraph. A "Supervisor Agent" routes natural language queries to specialized "Worker Agents" (e.g., Small Molecules, Polymers, Reactions) that perform self-reflective RAG loops to ensure relevance and accuracy.

Results and Discussion

A key finding of this study is the superiority of foundation model embeddings over traditional rule-based fingerprints in capturing semantic chemical similarity. As illustrated in Figure 4, MoLFormer-based metrics (Cosine and Euclidean distance) often identify structural analogues that traditional methods (Tanimoto, RDKit, MACC, Dice) fail to recognize.

Comparison of similarity metrics for small molecule sets A and B. Heatmaps show that foundation model-derived scores (Cosine, Euclidean) capture semantic relationships that traditional fingerprints (Tanimoto, RDKit) may miss.

For instance, in molecule set A, structure 1b exhibits a high Cosine similarity (0.93) to the query 1a, whereas the Tanimoto similarity is significantly lower (0.17). This suggests that the foundation model captures a deeper "chemist's intuition" regarding molecular topology that bitstring-based fingerprints overlook.

Furthermore, the multi-agent system demonstrated robust performance in complex research tasks. When queried about aromatic polyethers, the system successfully navigated a polymer vector store containing millions of entries, retrieved relevant analogues, and synthesized a technical report on their nucleophilic substitution pathways. The multi-modal component also showed high precision, retrieving chemical structures from NMR spectral images with an L2 distance as low as 0.0109 for top matches.

Conclusion

By shifting from text-centric to structure-centric RAG, we have demonstrated a significant advancement in the capabilities of autonomous chemical agents. The use of chemistry foundation models allows for more intuitive and effective information retrieval, mirroring the reasoning processes of human researchers. Future work will focus on expanding the model's token limits to accommodate larger macromolecules and integrating additional characterization modalities such as IR spectra and rheology data to create a truly universal AI lab assistant.

Reference: Park, N. H., et al. "Leveraging Chemistry Foundation Models to Facilitate Structure Focused Retrieval Augmented Generation in Multi-Agent Workflows for Catalyst and Materials Design." IBM Research–Almaden. (2024).

Summary updated

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.
BibTeX
@inproceedings{park2025leveraging,
  author = {Nathaniel H. Park and Tiffany J. Callahan and James L. Hedrick and Tim Erdmann and Sara Capponi},
  title = {{Leveraging Chemistry Foundation Models to Facilitate Structure Focused Retrieval Augmented Generation in Multi-Agent Workflows for Catalyst and Materials Design}},
  booktitle = {AI4Science @ NeurIPS},
  year = {2025},
  url = {https://openreview.net/forum?id=RGbK4xOO3c},
}
Figure