A research team led by Professor Joung Joonyoung of the Department of Chemistry at the Graduate School of Kookmin University (President Jeong Seung Ryul) has developed an artificial intelligence (AI) model that predicts the properties of mixtures based on the ratios of their constituent chemicals.

△ Photo: Overview of “MEMo,” an AI model that predicts the properties of mixtures containing various chemicals

Kim Ji-hwan, a student in the integrated master’s-doctoral program, and Kim Chan-joong, an undergraduate research student, participated in this study. The research findings were published in *Chemical Engineering Journal*, a leading SCI-indexed international journal in the field of chemical engineering, under the title “Encoding composition-dependent intermolecular interactions into learned molecular representations.” This journal is a top-tier (Q1) international journal in the JCR with an Impact Factor of 13.2.
 
Most chemical products around us are mixtures of multiple substances rather than a single substance. Typical examples include solvents used in factories, electrolytes in batteries, automotive fuels, and formulations for cosmetics and pharmaceuticals. The properties of a mixture depend not only on which substances it contains but also on the proportions of each substance, and these changes do not necessarily correlate linearly with the mixing ratios. Since there are limitations to experimentally measuring all possible combinations and ratios, the importance of AI technology capable of predicting the properties of mixtures without experimentation is growing.
 
Existing AI models primarily analyzed mixtures by first learning the characteristics of each molecule individually and then calculating an average based on the composition ratio. This is similar to estimating the final flavor of a dish by tasting each ingredient separately and calculating an average, without actually tasting the finished dish. This approach had limitations in learning the interactions between molecules and the changes resulting from different composition ratios.
 
To address these issues, Professor Joung’s research team developed a new AI model called “MEMo (Mixture Embedding via Mole Fraction).” MEMo is designed to receive input on which molecules are mixed and in what proportions from the very beginning of the process of understanding molecules. The front end of the model identifies the characteristics of each molecule, while the back end learns interactions as different molecules exchange information with one another. This enables the AI to recognize even the same molecule differently depending on which molecules it is mixed with and in what proportions.
 
Based on over 860,000 experimental data points from the U.S. National Institute of Standards and Technology (NIST), the research team constructed evaluation datasets for five properties—density, sound speed, vapor pressure, refractive index, and surface tension—and validated the model’s performance. As a result, MEMo demonstrated higher prediction accuracy than existing leading methods for four out of the five properties.
 
In particular, the performance difference was notable in mixtures with strong intermolecular interactions—such as ionic liquids or salt solutions—and in mixtures containing new molecules that the AI had not encountered during the training process. Furthermore, since a single model can handle everything from pure substances to mixtures of two or three substances, it can be utilized to predict the properties of a wide variety of chemical substances. The research team has made both the evaluation dataset and the model code publicly available so that other researchers can freely utilize them.
 
This study is significant in that it trained the AI to learn not only the characteristics of the molecules themselves but also the intermolecular interactions that vary depending on the mixing ratios when predicting the properties of mixtures. In particular, it is expected to reduce the experimental burden associated with various composition conditions and contribute to improving the efficiency of material development across multiple industrial sectors, such as battery electrolytes and eco-friendly solvents.
 
Professor Joung Joonyoung stated, “The properties of mixtures depend on ‘who’ the molecules interact with and ‘how much’ they interact; the core of this research is that we designed the AI to understand molecules based on the same principle,” adding “Since it enables rapid prediction of the properties of various mixtures without experimentation, we expect it to help reduce the time and cost of material development across a wide range of industries dealing with mixtures, including battery electrolyte design, the development of eco-friendly solvents, and the search for carbon dioxide capture materials,” he added.

△ Photo: (From left) Professor Joung Joonyoung of Kookmin University, Kim Ji-hwan, a student in the integrated master’s-doctoral program, and Kim Chan-joong, an undergraduate research assistant