Member and Video Highlight – Carnegie Mellon University

In a recent Nature Machine Intelligence commentary, Xin, Kitchin, and Kulik advocate for a new scientific paradigm—agentic science—defined by (semi-)autonomous AI agents that can reason, plan, and interact with both digital and physical environments. Building on advances in large language models (LLMs), multimodal learning, and automation, agentic AI systems are now being developed to autonomously manage entire research workflows: hypothesis generation, experimental design, data analysis, and even manuscript drafting. While promising, agentic science presents challenges including hallucination, prompt fragility, and a lack of standardized evaluation metrics—issues especially critical for chemical engineering workflows reliant on rigor, safety, and reproducibility. The authors call for standardized protocols, domain-specific benchmarks, and human-in-the-loop strategies to ensure that agentic AI enhances rather than undermines scientific integrity. This vision aligns with the mission of CACHE to advance computer-aided methods that augment chemical engineering discovery and education.

- Xin, H., Kitchin, J. R., & Kulik, H. J. (2025). Towards agentic science for advancing scientific discovery. Nature Machine Intelligence, 7, 1373-1375.
  https://doi.org/10.1038/s42256-025-01110-x