Predictions of High-Order Electric Properties of Molecules: Can We Benefit from Machine Learning?

There is an exigency of adopting machine learning techniques to screen and discover new materials which could address many societal and technological challenges. In this work, we follow this trend and employ machine learning to study (high-order) electric properties of organic compounds. The results of quantum-chemistry calculations of polarizability and first hyperpolarizability, obtained for more than 50,000 compounds, served as targets for machine learning-based predictions. The studied set of molecular structures encompasses organic push-pull molecules with variable linker lengths. Moreover, the diversified set of linkers, composed of alternating single/double and single/triple carbon- carbon bonds, was considered. This study demonstrates that the applied machine learning strategy allows us to obtain the correlation coefficients, between predicted and reference values of (hyper)polarizabilities, exceeding 0.9 on training, validation, and test set. However, in order to achieve such satisfactory predictive power, one needs to choose the training set appropriately, as the machine learning methods are very sensitive to the linker-type diversity in the training set, yielding catastrophic predictions in certain cases. Furthermore, the dependence of (hyper)polarizability on the length of spacers was studied in detail, allowing for explanation of the appreciably high accuracy of employed approaches.