Automating Chemistry With LLM-Based Agents

Explore the cutting-edge intersection of artificial intelligence and scientific research through this comprehensive seminar featuring two presentations on automating scientific discovery with large language models. Begin with Andrew White's primer on the automation of science, examining how exponential growth in research complexity and declining scientific productivity create intellectual bottlenecks that require automated solutions. Learn about scientific agents - large language models augmented with specialized tools - and their applications in automating literature research, molecular design, enzyme engineering, and bioinformatics analysis. Discover how these agents have been trained to exceed human performance across multiple scientific tasks through rigorous evaluation against expert researchers, and gain insights into the emerging field of scientific reasoning models. Continue with Sam Cox's detailed presentation on LLM-based chemistry agents, focusing on how these systems address the limitations of standard language models in complex chemistry problems by integrating expert-designed tools and updated knowledge sources. Examine ChemCrow, an LLM chemistry agent designed for organic synthesis, drug discovery, and materials design tasks, alongside MDCrow, which automates protein simulation setup, execution, and analysis using molecular dynamics tools. Understand how open-source models, when properly augmented with appropriate tools, can achieve performance comparable to frontier LLMs in chemistry applications, demonstrating the democratization potential of these technologies for scientific research automation.