Professor Joung Joon Young's research team from the Department of Applied Chemistry, majoring in Nanomaterials, at Kookmin University (President Jeong Seung Ryul) has jointly developed a next-generation artificial intelligence model for predicting chemical reactions, ‘FlowER (Flow Matching for Electron Redistribution)’, with Professor Connor W. Coley's team at the Massachusetts Institute of Technology (MIT). This model has been featured in (Connor W. Coley) at the Massachusetts Institute of Technology (MIT), has been published in the world's most prestigious academic journal, Nature, under the title ‘Electron flow matching for generative reaction mechanism prediction’.

Chemical reaction prediction is a core technology directly linked to various industries and daily life, including new drug development, eco-friendly material design, and energy storage technology. To create new materials, one must know which raw materials to react under what conditions to achieve the desired outcome. Incorrect predictions waste significant time and resources. Until now, this has primarily relied on chemists' experience and repeated experiments. However, complex reactions may require examining thousands of possibilities, making it practically impossible for humans to test every scenario. For this reason, AI capable of rapidly and accurately predicting complex chemical reactions is gaining attention as a tool that can dramatically enhance research efficiency.

The FlowER developed in this study predicts reactions step-by-step while strictly adhering to the ‘law of mass conservation,’ which previous AI models often overlooked. The research team redefined chemical reactions as ‘electron rearrangement’ problems and implemented a model capable of tracking intermediates by precisely conserving the number of electrons and atoms before and after the reaction. The team employed a method that maps how atoms connect within molecules and how electrons move, combining this with the latest AI learning technique called ‘flow matching’. Flow matching is a technology that naturally connects and predicts states changing over time, similar to stitching together multiple scenes to create a movie.  

Thanks to this approach, the model can predict with high accuracy the differences between multiple pathways that may vary depending on reaction conditions, byproducts that often appear in experiments, and even novel reactions not found in textbooks. Knowing these differences, byproducts, and new reactions in advance can reduce unnecessary trial and error, enabling the faster and safer production of desired substances.

A key advantage of FlowER is its ability to rapidly adapt to entirely new types of chemical reactions with very few examples. For instance, after being shown just 32 examples, it correctly identified the reaction pathway over 80% of the time for previously unseen reactions. Furthermore, the predicted reaction pathways can be verified against actual physical and chemical laws using quantum chemistry calculations, significantly reducing time and costs from the experimental design stage. This technology is expected to expand its application across diverse fields, including new drug development, eco-friendly catalyst design, and energy material research.

Professor Joung Joon Young of Kookmin University explained, “This research demonstrates that AI can go beyond merely providing ‘correct answers’ to infer reaction mechanisms in a manner akin to the language chemists use,” adding, “Notably, prediction performance improved significantly simply by designing the model to adhere to the law of conservation of mass, one of the most fundamental principles in science.” He further stated, “Going forward, we plan to expand our research beyond electrochemistry, atmospheric chemistry, and biochemical reactions to discover new chemical reactions not yet reported in the world.”

He further noted, “This research exemplifies the fusion of chemistry, a fundamental discipline, with artificial intelligence. It demonstrates that performance can be enhanced by incorporating fundamental laws that existing AI researchers often overlook in fields where domain knowledge is crucial.” He concluded, “This presents the potential for more precise application of AI across various scientific fields in the future.”