Domestic researchers have developed artificial intelligence (AI) technology that learns electronic level information to predict molecular properties. Electronic level information is fundamental information that represents the properties of substances. It is typically obtained through quantum mechanics calculations, which incur high expenses. The research team has enabled predictions of molecular properties using AI without incurring significant expenses.
The Korea Research Institute of Chemical Technology announced on the 15th that a joint research team led by senior researcher Na Kyung-seok and Professor Park Chan-young of KAIST has developed a self-supervised diffusion model-based molecular representation learning technology that precisely predicts material properties based on molecular electronic level information without high-cost quantum mechanics calculations. The technology was presented at the International Conference on Learning Representations (ICLR), an AI academic conference, last April.
The research team decomposed complex molecules into smaller ones. AI infers information using electronic level information from the smaller molecules. This allows for the prediction of the properties of the original molecules. The method is self-supervised, where AI learns independently without human intervention.
This technology recorded world-class accuracy in predicting molecular properties based on 29,711 experimental data points. For the optical properties of molecules, the prediction accuracy was 88%. This was higher than the prediction accuracy of existing AIs, which was between 31% and 44%.
The research team explains that the technology can be utilized in industries requiring new material development. They noted, "It can be used in the development of new materials in the semiconductor, display, and pharmaceutical sectors," adding that it can reduce the expense of purchasing computer equipment for quantum mechanics calculations.