A charge density prediction model for organic molecules using Deep Neural Networks

In an attempt to accelerate the pace of materials discovery, the community is increasingly using machine learning (ML) based techniques to rapidly map structure property relationships in materials. The Machine Learning methods has the advantage that their computational costs are negligible compared to those of the customary methods like density functional theory (DFT). The simulation properties of polymers is one such example where this limitation is most apparent . Here we propose a method to accurately predict charge densities of large organic systems by learning from pre-calculated examples of smaller systems. We present a novel fingerprint scheme which can numerically represent the local atomic environment. We then use a Neural Networks based models to learn the functional map between the local environments and the charge densities. Further, we introduce a recursive approach to improve these models by selectively incorporating poorly predicted examples. These charge densities are then used to calculate properties which depend on the charge density distribution like the dipole moments, dielectric properties for a host of unseen molecules thus eliminating the need to explicitly solve the laborious Kohn-Sham equations.