Developing a new generation of AI systems designed to understand chemistry through real scientific principles rather than pattern matching alone. Led by Connor Coley, the work focuses on building AI models that respect physical laws such as conservation of mass and electron behavior while predicting chemical reactions and designing molecules. Scientists believe this approach could make AI far more reliable for drug discovery, materials science, and industrial chemistry.
One major problem with many existing AI chemistry systems is that they often behave like large language models trained only on statistical patterns. Researchers say these systems can sometimes generate chemically impossible reactions by effectively “creating” or “deleting” atoms during predictions. Coley’s team is trying to solve this by embedding chemical reasoning directly into model architecture. Their newer system, called FlowER, explicitly tracks electron movement during reactions so predictions remain physically consistent and scientifically valid.
The broader goal is to create AI systems capable of accelerating molecule discovery while remaining grounded in real-world chemistry. Coley’s research group combines machine learning, cheminformatics, and laboratory automation to help design new medicines, catalysts, and materials more efficiently. Projects within the lab explore generative AI for synthesizable molecules, reaction prediction, retrosynthesis planning, and autonomous experimentation systems that can propose and test chemical hypotheses with minimal human intervention.
Researchers believe physically grounded AI models could eventually transform scientific discovery by reducing the enormous time and cost involved in chemical experimentation. Instead of relying entirely on trial and error, future systems may help scientists predict viable compounds, identify reaction pathways, and automate parts of laboratory research with much higher reliability. However, Coley and his collaborators also acknowledge that current systems still have limitations, especially involving complex catalytic chemistry and rare reaction types. Even so, experts increasingly see “AI for Science” as one of the most promising long-term applications of artificial intelligence beyond consumer chatbots and automation.
