Prediction of electronic density of states in guanine-TiO2 adsorption model based on machine learning

Abstract

The electronic density of states is a property of the material that is extensively used in quantum systems in condensed matter physics. It refers to the energy level of the electrons in a solid crystal. One of the most current ways to compute it is by Density Functional Tight Binding (DFTB), given the geometry of the material. Nevertheless, this computation could be very computationally demanding, although applied to some materials with a very reduced number of atoms. This paper presents a method to deduce the electronic density of states, which is based on a neural network, thus, almost linear with respect to the number of atoms in the material. Specifically, we have applied our method to a metal oxide structure interacting with a nucleic base guanine. We have focused on stoichiometric and O-defective anatase TiO2 (101) surfaces. The data set of the electronic density of states needed to train the neural network has been obtained by a DFTB+ numerical solver given an initial molecular geometry model, which computed a track of time-dependent geometry and their associated electronic density of states. We validated that the predicted electronic density of states deduced by our method and by DFTB tends to be very similar, thus, opening the door to other computations such that introducing our method in the process of generating the time-dependent geometry analysis.