Incorporation of a process that applies physical theory to perform calculations into deep learning, and clarification of the learning process
The newly developed technique features prediction of the physical properties of molecules by expressing and transiting both a wave function and the electron density within the deep learning model, based on density functional theory. This eliminates the black box nature of the model and increases reliability by enabling researchers and engineers to verify and interpret the prediction results when applied to material development and drug discovery.
The process and result of the technique for learning and predicting the physical property E of a molecule from its structure are as follows: Information on the molecular structure (input) is converted into the atomic wave function φ, then the wave function ψ and electron density ρ of the molecule are calculated, and finally the physical property is output. This model is constructed using the neural networks, and is trained and performs prediction using a large-scale database of molecular structure and physical properties. The predictions achieve the accuracy demanded in theoretical calculations, and predictions for tens of thousands of molecule types can be made in a matter of minutes, enabling to increase the speed of theoretical calculations by a factor of 100,000 times or more.