Improving Quantum Mechanical Calculations Using Graph Neural Networks to Predict Energies from Atomic Structures

Density functional theory (DFT), a popular quantum-mechanical method for electronic structure calculations, has been utilized by computational chemists for decades to calculate various properties of a wide range of chemicals and materials. The DFT approach, which uses functionals to map a chemical structure’s electron density to its energy, is less computationally expensive than wavefunction based ab inito calculation methods, while still maintaining a high degree of accuracy. However, DFT calculations are quite slow which limits their application in molecular dynamics (MD) simulations and large-scale systems. Machine learning (ML) offers the potential to learn the electron density to energy functional from DFT data, eliminating the need to solve the computationally expensive Kohn-Sham equations, and thus allowing for calculations of much larger atomic systems and time scales. This would enable large scale MD simulations with DFT-level accuracy, a better alternative to the less-accurate empirical-based MD simulations which are currently used to model large scale atomic systems. In the work, we use graph neural networks (GNNs) to predict electron densities and energies directly from atomic structure. We examine the efficacy of the GNN models through electron densities, energies and forces on a set of 3D structures from small organic molecules of under 30 atoms.