Mayur Singh1,Matthew Barry1,Satish Kumar1
Georgia Institute of Technology1
Mayur Singh1,Matthew Barry1,Satish Kumar1
Georgia Institute of Technology1
Prediction of dynamical properties of materials through Molecular Dynamics (MD) simulations comes with the established problems such as exponential increasing computational cost with increasing size, limited time scale, and inaccuracy of results from using empirically derived interatomic potentials compared to first principles Density Functional Theory (DFT) simulations. Neural Network Interatomic Potentials (NNIPs) have lower computational cost and comparable accuracy to the first principles calculations. However, one problem faced by NNIPs is the creation of mostly simplified or expert-guided <i>ad hoc</i> selection of the salient material structure descriptors to describe interatomic forces, when in actuality these interactions happen on complex 3D atomic structures. In this work, we introduce the Voxelized Atomic Structure (VASt) potential for Molecular Dynamics. VASt is a framework for creating interatomic potentials through the voxelization of the atoms in an atomic structure. This creates a 3D representation of the structure, which can be trained in a Convolutional Neural Network with 3D convolutions. We use the VASt potential in MD simulations to calculate the mass diffusion coefficient of two-component systems. Performance of many metallic alloys are highly dependent on the mass diffusion coefficient of the components, notably Ti-Al alloys. We consider a 2-component system of Al and -Ti and study the diffusion of Al into -Ti in a wide range of temperature, i.e., 300-1500 K. Overall, our approach has great potential to predict diffusive, mechanical and thermal properties of multi-component systems.