Foundation Models and Multi-Agent Systems for Polymer and Catalyst Design

Abstract

The development of foundation models which can generalize across prediction tasks is imperative for the construction of seamless workflows to enable AI-guided design of novel polymeric materials. Here, we will discuss our efforts in addressing the many challenges of building an effective foundation model for polymers, including strategies for efficient representation of complex materials and construction of relevant benchmarking datasets. In addition to the foundation models themselves, facilitating straightforward interactions between the model and human researchers is a critical aspect in realizing their utility within the experimental research workflows. In this regard, we will discuss how materials foundation models can be incorporated within LLM-powered, multi-agent systems to facilitate not only property prediction but also multimodal retrieval augmented generation (RAG) tasks to provide grounded and salient answers to researcher inquiries. Importantly, these RAG tasks will demonstrate how the combination of chemistry foundation models with LLMs can enable highly challenging queries for polymer topology or leveraging characterization data can be performed. Finally, the application of these technologies for development of sustainable materials will also be discussed.