Machine learning algorithm predicts density of states in organic molecules

Researchers at the Institute of Industrial Science, The University of Tokyo, have made significant progress in the field of organic chemistry by developing a machine-learning algorithm capable of predicting the density of states within organic molecules. This breakthrough has the potential to revolutionize the analysis of carbon-based molecules, aiding organic chemists and materials scientists in their investigations.

The density of states, which refers to the number of energy levels available for electrons to occupy within a material’s molecules in their ground state, is crucial for understanding the chemical properties of a substance. Traditionally, experimental techniques like core-loss spectroscopy have been employed to determine the density of states. However, interpreting the results of these methods, particularly in the case of core-loss spectroscopy, can be challenging.

Core-loss spectroscopy combines two techniques: energy loss near-edge spectroscopy (ELNES) and X-ray absorption near-edge structure (XANES). By irradiating a material sample with a beam of electrons or X-rays, scientists can measure the scatter of electrons and the energy emitted by the molecules, enabling them to assess the density of states within the molecules. However, this information is only available for the unoccupied (electron absent) states of the excited molecules.

To overcome this limitation, the team at the Institute of Industrial Science trained a neural network machine-learning model to analyze core-loss spectroscopy data and predict the density of electronic states. They created a comprehensive database by calculating densities of states and corresponding core-loss spectra for over 22,000 molecules, incorporating simulated noise. Subsequently, the algorithm was trained on core-loss spectra and fine-tuned to accurately predict the density of states for both occupied and unoccupied states in the ground state.

Lead author Po-Yen Chen explains, “We attempted to extrapolate predictions for larger molecules using a model trained on smaller molecules. We discovered that the accuracy can be improved by excluding tiny molecules.” This insight proved valuable in enhancing the algorithm’s performance.

Moreover, the team discovered that by applying smoothing preprocessing techniques and introducing specific noise to the data, they could further enhance the predictions of density of states. This improvement brings the prediction model closer to real-world applications, enabling researchers to gain deeper insights into the material properties of molecules and expedite the design of functional compounds, including pharmaceuticals and other exciting substances.

Senior author Teruyasu Mizoguchi emphasizes the significance of their work, stating, “Our research can contribute to a better understanding of material properties and accelerate the development of functional molecules.” The implications of this advancement extend beyond organic chemistry, reaching into various domains where carbon-based molecules play a vital role, such as OLED displays and other technologies of the present and future.

Source: University of Tokyo

Leave a Comment