Do not go gentle into that good night, old age should burn and rave at close of day; Rage, rage against the dying of the light, though wise men at their end know dark is right, because their words had forked no lightning they, do not go gentle into that good night.
Kookmin University professor’s research on AI models published in Nature
An imaginative rendering of Kookmin University applied chemistry research team's FlowER model / Courtesy of Kookmin University
By Jhoo Dong-chan
A research team led by assistant professor Jung Jun-young from the department of applied chemistry at Kookmin University has developed a next-generation artificial intelligence (AI) model for chemical reaction prediction, together with professor Conner W. Coley’s group at Massachusetts Institute of Technology (MIT).
Their collaborative model, named FlowER (Flow Matching for Electron Redistribution), was published in science journal Nature in an article titled “Electron flow matching for generative reaction mechanism prediction,” a Kookmin University official said in a press release.
The Nature article noted the model’s potential applications in various industries, including drug discovery, sustainable materials and energy storage, thanks to its ability to accurately predict chemical reactions. Traditionally, researchers relied heavily on expertise and repeated experiments, but complex reactions can involve thousands of variables, making manual prediction impractical. AI models that can rapidly and precisely forecast reaction pathways are therefore considered transformative tools.
The FlowER model addresses a key limitation of existing AI systems by strictly adhering to the law of mass conservation while predicting reactions step by step.
The research team redefined chemical reactions as problems of “electron redistribution,” ensuring conservation of atoms and electrons and enabling the model to trace intermediate states.
By mapping atomic connectivity and electron flows as a reaction map and combining it with a state-of-the-art training technique called flow matching, the model can predict reactions with high accuracy.
FlowER can also distinguish between multiple reaction pathways, byproducts and even previously unreported reactions, according to the article. It has demonstrated high adaptability to new reaction types while requiring only a small number of examples. In tests, the model correctly identified reaction pathways over 80 percent of the time after being trained on just 32 examples.
The predictions can also be validated with quantum chemistry calculations, reducing trial-and-error in lab experiments.
“The research highlights how incorporating fundamental scientific principles into AI design can significantly strengthen performance,” Jung said.
“We expect FlowER to be applied in drug development, catalyst design, energy materials and the discovery of entirely new chemical reactions.”
Kookmin University applied chemistry assistant professor Jung Jun-young / Courtesy of Kookmin University