Optimization of molecular characteristics via machine learning based on continuous representation of molecules

We demonstrate an automatic materials design method using continuous representation of molecule and its atomic arrangement via a neural network algorithm. This method is applied to optimizing and predicting the HOMO-LUMO gap within the molecules composed of carbon, oxygen, nitrogen, fluorine, and hydrogen. Adopting the Quantum Machine 9 (QM9) dataset as a training dataset for the molecules, we first established a continuous representation of molecules in a latent space, then predicted molecules that have target values of the HOMO-LUMO gap. In the gap maximization calculation, CF4 with the largest gap value in the QM9 dataset was automatically found despite there is no a priori data for the gap. In the case of a target gap value of 0.10 hartree, we found a new molecule whose gap value is closer to 0.10 hartree than any other molecules in the QM9 dataset.