Deep-learning molecular dynamics simulation of pressure-driven transformation for bulk TiO2

Accurate interatomic interaction potentials on the basis of deep-learning methods can provide new opportunities for molecular dynamics (MD) simulations on large spatial and temporal scales in the field of material phase transformation (PT). Here, we obtained a new interatomic interaction potential for bulk TiO2 through deep-learning method trained on our theoretical calculations with density functional theory (DFT). Our MD simulations with deep-learning potential (DP) revealed that a large number of chiral quarter vortices emerge in the large bulk TiO2 during martensitic-like columbite -> baddeleyite and orthorhombic CaCl2-type phase -> baddeleyite PTs. Ti and O atoms move in a collective and synchronous order and then lead to Bain distortion in the central area between four different vortices. A certain shear stress (theoretically around 5.2 GPa) has to be provided to drive successfully the reverse PT of columbite -> baddeleyite phase. The intermediate phase with orthorhombic CaCl2-type symmetry (space group No. 58, Pnnm) appears under hydrostatic pressure with 7.6 GPa for rutile -> baddeleyite PT. The PT path for intermediate phase -> baddeleyite PT is nearly the same as that of columbite -> baddeleyite PT. Our MD simulations with DP can provide a new understanding for martensitic-like PT behavior in large bulk materials and the ghostly existence of intermediate phase since 1971.