A charge density prediction model for hydrocarbons using deep neural networks

The electronic charge density distribution ρ(r) of a given material is among the most fundamental quantities in quantum simulations from which many large scale properties and observables can be calculated. Conventionally, ρ(r) is obtained using Kohn–Sham density functional theory (KS-DFT) based methods. But, the high computational cost of KS-DFT renders it intractable for systems involving thousands/millions of atoms. Thus, recently there has been efforts to bypass expensive KS equations, and directly predict ρ(r) using machine learning (ML) based methods. Here, we build upon one such scheme to create a robust and reliable ρ(r) prediction model for a diverse set of hydrocarbons, involving huge chemical and morphological complexity /(saturated, unsaturated molecules, cyclo-groups and amorphous and semi-crystalline polymers). We utilize a grid-based fingerprint to capture the atomic neighborhood around an arbitrary point in space, and map it to the reference ρ(r) obtained from standard DFT calculations at that point. Owing to the grid-based learning, dataset sizes exceed billions of points, which is trained using deep neural networks in conjunction with a incremental learning based approach. The accuracy and transferability of the ML approach is demonstrated on not only a diverse test set, but also on a completely unseen system of polystyrene under different strains. Finally, we note that the general approach adopted here could be easily extended to other material systems, and can be used for quick and accurate determination of ρ(r) for DFT charge density initialization, computing dipole or quadrupole, and other observables for which reliable density functional are known.