Theoretical Study On Simulation For The Diffusion Of Oxygen Vacancy In Titanium Dioxide Based On Deep Learning

Energy is the material basis for the survival and development of human society,traditional fossil energy is non-renewable and generates a lot of polluting gases from combustion.Solar energy is a widespread,valuable and new energy which attracts great attentions.Photocatalysis technology,one of the effective ways to use solar energy,has been widely used in the fields of health care,wastewater degradation,highway environmental protection,atmospheric purification,etc.Titanium dioxide（Ti O₂）is one of the most used photocatalysts with the advantages of high stability,non-toxicity,low price and good catalytic activity,etc.The band gap of Ti O₂ rutile phase（Rutile）is about 3.0 e V and the visible light utilization rate is less than 5%.Research literature found that oxygen vacancy defect（Ov）can effectively regulate the material energy band structure and promote carrier separation;as a photocatalytic active center,it has a significant impact on promoting the catalyst interfacial reaction and enhancing the photocatalytic efficiency.Numerous studies have found that the diffusion of Ov is a common phenomenon in catalytic reactions;therefore,it is of great significance to study and simulate the diffusion mechanism of Rutile oxygen vacancies for the study of Ti O2photocatalysis.After fully considering the advantages and disadvantages of ab initio molecular dynamics（AIMD）and classical molecular dynamics,this paper uses deep potential molecular dynamics method（DPMD）based on the deep potential（DP）model which is widely used in the field of computational materials to simulate and study the diffusion mechanism of oxygen vacancies.The advantage of DPMD is that it has the computational accuracy of first principles while having computational efficiency several orders of magnitude higher than that of AIMD.In this paper,we focus on the Ti O₂ rutile phase,where the most stable（110）surface is used as the main study model.The initial training set is firstly calculated from the simulation trajectories of AIMD with different Ov at 330 K.After several iterations using the DPMD method combined with the enhanced sampling technique,we got the final converged DP model which is applicable to the temperature interval 330-700 K.In order to verify the accuracy of the DP model,a series of comparisons were performed between using the DPMD method and density functional theory（DFT）calculations for properties such as diffusion barrier values,oxygen vacancy formation energy,interatomic force deviation,and total energy of the system,respectively.It shows that the results obtained by the two methods agree well,indicating that the DP model has the computational accuracy of DFT.The comparison of the two methods also illustrates that the DP potential describes the structures that are not in the training set well,which means that the DP potential is scalable,thanks to the rotation,translation and substitution invariance guaranteed during the DP potential training.In this paper,we perform simulations on time scales up to nanoseconds（ns）for a variety of Ov structures and explore the diffusion processes of a variety of Ov and the main factors affecting the diffusion potential and diffusion behavior of Ov.We also plot the free energy surface at 330 K with respect to the diffusion process of the main oxygen vacancy（bridge Ov）along the[001]direction in the Rutile system.The main conclusions of this paper are:Rutile（110）Ov is more stable in nature at room temperature and less prone to diffusion;the stability of Ov varies between surface and bulk phases,and Ov diffusion from the（110）surface to the bulk phase is difficult,while the opposite direction is easier.Finally,this paper briefly illustrates the relationship between the efficiency of DPMD and the number of atoms in the calculated system.This study illustrates that DPMD as an artificial intelligence（AI）tool has a very promising application in simulating material-related properties at the atomic scale.