TensorAlloy: An automatic atomistic neural network program for alloys
Description
Atomistic modeling is important for studying physical and chemical properties of materials. Recently, machine learning interaction potentials have gained much more attentions as they can provide density functional theory level predictions within negligible time. The symmetry function descriptor based atomistic neural network is the most widely used model for modeling alloys. To precisely describe complex potential energy surfaces, integrating advanced metrics, such as force or virial stress, into training can be of great help. In this work, we propose a virtual-atom approach to model the total energy of symmetry function descriptors based atomistic neural network. Our approach creates the computation graph directly from atomic positions. Thus, the derivations of forces and virial can be handled by TensorFlow automatically and efficiently. The virtual atom approach with AutoGrad within TensorFlow allows for efficient training to not just energies and forces, but also virial stress. This new approach is implemented in our open-source program TensorAlloy, which supports constructing machine learning interaction potentials for both molecules and solids. The QM7 and SNAP/Ni–Mo datasets are used to demonstrate the performances of our program.