Researchers at the Institute of Industrial Science, The University of Tokyo, have developed a machine-learning algorithm that predicts electron energy levels in organic molecules, utilizing a database of over 22,000 compounds. This advancement is significant for the fields of organic chemistry and materials science, which are essential for technologies like organic light-emitting diodes (OLEDs) and pharmaceuticals. The study focuses on predicting the density of electronic states, which represents the energy levels available for electrons in a molecule. Traditional methods, such as core-loss spectroscopy, are complex and often challenging to interpret, especially because they typically provide data only on unoccupied states of excited molecules.
To enhance understanding of these structures, the research team trained a neural network model on core-loss spectroscopy data to accurately predict both occupied and unoccupied energy states. The model’s accuracy was improved by focusing on larger molecules and employing noise simulation techniques. Lead author Po-Yen Chen noted that excluding smaller molecules improved prediction accuracy. Senior author Teruyasu Mizoguchi emphasized the potential of this approach to aid in the understanding of material properties, ultimately accelerating the design of new functional molecules, including drug compounds.
