A Kookmin University team has developed MEMo, an AI model that predicts the properties of chemical mixtures by learning how their molecules interact and in what proportions. Tested against more than 860,000 experimental measurements, it outperformed conventional methods on four of five properties, according to The Korea Times.
Watch Desk analysis
What happened
The model, Mixture Embedding via Mole Fraction, first learns characteristics of individual molecules, then lets them exchange information so it can account for interactions between substances. That is a different approach from estimating a mixture’s behaviour by simply averaging the properties of its ingredients.
The team evaluated predictions for density, speed of sound, vapour pressure, refractive index and surface tension. The university researchers reported particularly strong improvements for mixtures with strong molecular interactions, including ionic liquids and aqueous salt solutions, and for mixtures containing molecules absent from the training data. The findings were published in the Chemical Engineering Journal.
Why it matters
Mixtures are everywhere from battery electrolytes and industrial solvents to fuels, cosmetics and medicines. Predicting how a formulation behaves without having to make and test every version could help researchers narrow the field before they reach for the lab equipment. The team says possible applications include battery materials, eco-friendly solvents and materials for capturing carbon dioxide.
The result is promising, but not a licence to skip the laboratory altogether. MEMo beat conventional methods on four of the five tested properties; that leaves one where it did not, and the available account does not give the detailed comparative scores. A useful model can guide experiments without making them disappear by magic.
Our read
This is a concrete research advance with a practical target: better predictions for mixtures, where changing proportions can change how molecules behave. The most interesting signal is its reported performance on unfamiliar molecules, though further detail and testing will show how reliably that advantage carries into real materials work.
What to watch
- Detailed results for each of the five properties, including the test where MEMo did not outperform conventional methods.
- Whether independent researchers reproduce the gains, especially on mixtures containing unfamiliar molecules.
- Evidence that the model improves the speed or cost of developing useful materials.
Discussion spark: Would you trust an AI model to narrow down which material formulations deserve lab testing, or should experimental evidence remain the gatekeeper before any promising candidate moves forward?
Sources and evidence
- Source update (8 November 2026, 02:56 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.