Organic chemistry, the study of carbon-based molecules, is essential for many present and future technologies, including organic light-emitting diode (OLED) displays. It is the foundation of the science of living beings. Predicting a material’s chemical properties requires knowledge of the molecules’ electronic structures.
A machine-learning algorithm was developed to predict the density of states within an organic molecule, or the number of energy levels that electrons can occupy in the ground state within the molecules of a material, in a study that was recently published in The Journal of Physical Chemistry by researchers at the Institute of Industrial Science, The University of Tokyo. When analyzing carbon-based molecules, organic chemists and materials scientists can greatly benefit from these predictions, which are based on spectral data.
It might be challenging to comprehend the experimental methods that are frequently employed to determine the density of states. This is especially true for the technique called core-loss spectroscopy, which combines X-ray absorption near-edge structure (XANES) and energy loss near-edge spectroscopy (ELNES). These techniques expose a sample of material to an electron or X-ray beam; the scattering of electrons that results and measurements of the energy emitted by the material’s molecules allow for the measurement of the density of states for the target molecule. The electron missing (unoccupied) states of the excited molecules are the sole states in which the spectrum contains information.
To solve this problem, the researchers at The University of Tokyo’s Institute of Industrial Science created a neural network machine-learning model to examine the data from core-loss spectroscopy and forecast the density of electronic states. The densities of states and accompanying core-loss spectra for more than 22,000 molecules were first calculated to create a database. They also included some fake noise. The system was then trained on core-loss spectra and optimized to forecast the correct state density at the ground state, including both occupied and unoccupied states.
Using a model trained on smaller molecules, we attempted to extrapolate predictions for larger molecules. We found that eliminating small compounds can increase accuracy, says lead author Po-Yen Chen.
The researchers also discovered that by utilizing preprocessing for smoothing and adding certain noise to the data, it is possible to improve density of state predictions, which can hasten the adoption of the prediction model for usage on actual data.
Senior author Teruyasu Mizoguchi says, their study can facilitate the design of functional molecules and aid researchers in understanding the material properties of molecules. This can apply to pharmaceuticals and other fascinating substances.