(173an) Deepcdp: Deep Learning Charge Density Prediction

Many applications in materials chemistry require information about charges and electron densities in materials and how they respond dynamically to chemical reactions, ion transport, etc. First-principles techniques such as density functional theory (DFT) provides both information about the atom positions and electron densities. This has allowed extensive research at the electron level and has enabled major developments in the understanding of materials chemistry. However, DFT can only be used for relatively small system sizes and short dynamic time scales because of the computational complexity of the calculations. To overcome this limitation, we have developed a local deep learning formalism for predicting charge densities only using atomic positions for molecules and condensed phase (periodic) systems. Our method, called DeepCDP (Deep-learning Charge Density Prediction), can predict electron densities for arbitrarily large systems by training it on DFT electron densities for small systems. We used weighted-Smooth Overlap of Atomic Orbitals (w-SOAP) to fingerprint atomic environments to their corresponding electron densities on a grid-point basis. We discuss techniques to determine optimized weighting and SOAP parameters that can effectively generate unique fingerprints depending on the material in consideration. Our DeepCDP formalism was tested for both molecular systems, like water, and periodic systems, like functionalized graphane. The models that were trained on small system sizes (<10 atoms) achieved electron density prediction accuracies of >99% on larger system sizes (>900 atoms). We demonstrate that DeepCDP achieves similar prediction accuracies for both charged and uncharged systems. We show how one can use the predicted densities for charged systems to perform interesting analysis such as computing the diffusion of charge centers over a large sheet of functionalized graphane. We make use of the ability of DeepCDP to predict densities locally to tackle intrinsic memory inefficiencies that occur during the evaluation phase, thus allowing us to perform rapid predictions for arbitrarily large system sizes.