Siya Zhu1,Axel van de Walle1,Jan Schroers2,Stefano Curtarolo3,Hagen Eckert3
Brown University1,Yale University2,Duke University3
Siya Zhu1,Axel van de Walle1,Jan Schroers2,Stefano Curtarolo3,Hagen Eckert3
Brown University1,Yale University2,Duke University3
The atomic-level structure of bulk metallic glasses (BMGs) is a key determinant of their properties. A large training dataset is required to construct a machine-learning modeling for the structure and properties of BMGs. However,<b> </b>an accurate representation of amorphous systems in computational studies has traditionally required large supercells that are unfortunately computationally demanding to handle using the most accurate ab initio calculations. To address this, we propose to specifically design small-cell structures that best reproduce the local geometric descriptors (e.g., pairwise distances or bond angle distributions) of a large-cell simulation. We rely on molecular dynamics (MD) driven by empirical potentials to generate the target descriptors, while we use reverse Monte Carlo (RMC) methods to optimize the small-cell structure. The latter can then be used to determine mechanical and electronic properties using more accurate electronic structure calculations. With high efficiency and accuracy of modeling and calculating properties of BMGs, a large and reliable dataset for ML of BMGs can be generated using our method. The method is implemented in the Metallic Amorphous Structures Toolkit (MAST) software package.