Kookmin University's MEMo AI Predicts Mixture Properties, Beating Conventional Models

A Kookmin University research team has built an AI model that predicts the properties of chemical mixtures more accurately than conventional methods, winning on four of five benchmark properties. The team validated it against more than 860,000 experimental measurements from the U.S. National Institute of Standards and Technology (NIST). Chemistry professor Joung Joon-young led the work, and the findings appeared in the Chemical Engineering Journal under the title "Encoding composition-dependent intermolecular interactions into learned molecular representations."

The model, called Mixture Embedding via Mole Fraction, or MEMo, matters because most chemical products people use every day are not pure compounds. They're mixtures. Industrial solvents, battery electrolytes, automotive fuels, and cosmetic and pharmaceutical formulations all blend multiple substances, and their behavior depends on exactly how much of each component goes in.

What problem does MEMo solve?

Predicting how a mixture will behave has long been harder than it sounds. The properties of a blend depend not only on which substances it contains but also on the proportions in which they're combined. And those changes don't scale in a simple, linear way with the mixing ratios. Double one ingredient and you don't necessarily get double the effect.

Conventional AI models have handled this clumsily. They typically learn the properties of individual molecules first, then calculate a weighted average based on each component's share of the mixture. That approach has a structural blind spot: it can't fully capture how molecules interact with one another, or how those interactions shift as the composition changes.

MEMo attacks that gap directly. The design, according to the university, folds in both the identity of each molecule and its mixing proportion from the earliest stages of molecular analysis. In the initial stages, the model identifies the characteristics of individual molecules. In later stages, it lets different molecules exchange information, so the model learns how they interact.

That architecture teaches the model a chemist's intuition in miniature. A molecule's behavior can vary depending on which other molecules surround it and in what proportions. MEMo recognizes that, rather than treating each ingredient as an isolated actor averaged into a blend.

How was it tested?

The evaluation was large and grounded in real lab data. The team drew on more than 860,000 experimental measurements from NIST, the U.S. standards agency, and tested predictive performance across five properties:

• density
• speed of sound
• vapor pressure
• refractive index
• surface tension

MEMo outperformed conventional methods on four of the five. The university's announcement did not specify which property the model failed to win, but the overall record marks a clear advance over the weighted-average approaches that dominate the field.

Two results stood out beyond the headline score. The improvement was particularly pronounced for mixtures with strong intermolecular interactions, such as ionic liquids and aqueous salt solutions. These are precisely the systems where molecule-to-molecule effects run strongest, and where simple averaging fails hardest.

MEMo also performed well on mixtures containing molecules it had never encountered during training. That generalization matters for practical use. Real-world materials development constantly involves new compounds, and a model that collapses outside its training data would be of limited value in an industrial lab.

Who built it, and what did they say?

The team sits in Kookmin University's Department of Chemistry at the Graduate School. Joung Joon-young, the professor leading the group, worked with two students: Kim Ji-hwan, enrolled in the university's integrated master's-and-doctoral program, and Kim Chan-joong, an undergraduate. The inclusion of an undergraduate author on a peer-reviewed Chemical Engineering Journal paper is notable in itself, though the university's announcement did not elaborate on individual contributions.

Joung framed the research in chemical terms rather than machine-learning jargon.

"The properties of a mixture depend on which molecules interact with one another and in what proportions," Joung said. "The key to this research was developing an AI model that understands molecular behavior based on the same principle."

In other words, the model's architecture mirrors the physics of the problem. Molecules influence each other, and their influence shifts with composition. A model built to represent that exchange of information outperforms one that treats components as independent inputs.

Why does this matter for materials development?

The economic stakes are straightforward. Materials development today often requires physical experiments: mix a candidate formulation, measure its properties, adjust, repeat. Each cycle costs time and money. Joung expects MEMo to cut both by rapidly predicting the properties of mixtures at different compositions without requiring lab work at every step.

That shifts the workflow from trial-and-error to computational screening. Instead of testing hundreds of electrolyte formulations in flasks, researchers could rank candidates in software and synthesize only the most promising ones. The university's announcement, made Thursday, did not claim MEMo eliminates experiments entirely, only that it reduces how many are needed.

The application areas the team named are commercially and politically significant:

• battery electrolyte design, a bottleneck area for energy storage and electric vehicles
• eco-friendly solvents, where regulators and customers push to replace toxic industrial chemicals
• carbon dioxide capture materials, a cornerstone technology for climate mitigation

All three fields involve exactly the kind of strongly interacting mixtures where MEMo showed its largest gains. Ionic liquids, for instance, appear repeatedly in carbon capture and advanced electrolyte research. Aqueous salt solutions are foundational to both battery chemistry and solvent work. The benchmark categories where the model excels overlap with the industries most eager for faster materials discovery.

What's the bigger picture?

The result fits a broader trend in AI for science: models that respect physical structure are beating generic ones. MEMo doesn't just average molecular fingerprints, as prior mixture models did. It encodes composition-dependent intermolecular interactions into learned molecular representations, as the paper's title puts it. That design choice, validated on 860,000-plus NIST measurements, delivered measurable accuracy gains.

For South Korean research, the work also signals a growing domestic capability in AI-driven chemistry. Kookmin University's team published in the Chemical Engineering Journal, a peer-reviewed international venue, with a small group that included students at early career stages.

The immediate next question is deployment. Joung said the model could find applications across fields involving mixtures, but the announcement did not detail industrial partnerships, licensing plans, or releases of the model's code. If MEMo or its successors become available to battery makers and solvent developers, the 860,000-measurement validation could mark the point where AI mixture prediction moved from an academic benchmark toward a standard tool in the materials lab.