Department of Chemistry - PhD Student Seminar - Closed-Loop Scientific Discovery in Chemistry with Agentic AI

Student: Mr. Jingyuan ZHU

Department: Department of Chemistry, HKUST

Supervisor: Prof. Haibin SU

Abstract

Agentic AI is emerging as a powerful paradigm for enabling closed-loop scientific discovery in chemistry. Traditional chemical research is often slowed by fragmented knowledge, vast experimental spaces, low throughput, heterogeneous data, interpretive bias, and irreproducible protocols. By integrating large language models with tool use, memory, planning, and action, agentic AI systems can propose hypotheses, design experiments, execute workflows, analyze results, and iteratively refine strategies. This field builds on the evolution of chemistry automation, from manual experimentation and rule-based retrosynthesis to flow chemistry, high-throughput experimentation, Bayesian optimization, and robotic self-driving laboratories. It also draws on the core architecture of chemical agents, including an LLM-based reasoning brain, standardized protocols that connect tools and humans, memory systems for persistent context, and self-evolving mechanisms that improve reusable skills through feedback. Representative applications span both physical wet-lab agents and digital in-silico agents, including autonomous reaction optimization, robotic synthesis platforms, quantum chemistry workflow agents, and AI-guided materials screening. Despite rapid progress, major challenges remain, such as hardware cost, hallucination, fragmented chemical knowledge, limited multimodal and 3D molecular understanding, and the lack of long-term group-level memory. Overall, agentic AI is moving chemistry toward an integrated, machine-executable, and self-improving discovery loop, while fully autonomous and reliable AI chemists remain an open frontier.