GPU-Algorithms for Interatomic Potentials Based on Machine learning

The accurate calculation of interatomic potentials plays a crucial role in atomistic modeling. The only accuracy often is not enough in real­world applications. If the number of atoms in atomistic systems rises, the accurate methods of the calculation potentials will not be computationally effective. Hence, machine learning interatomic potentials are a new class of potentials that can help paradigm shift in atomistic modeling. This thesis describes an atomistic modeling approach that relies on a special type of machine learning interatomic potential — moment tensor potentials (MTPs) — to approximate the quantum mechanical potential energy surface using nonlinear functions of atomic environment descriptors. In conjunction with GPU algorithm to improve inference of MLIP. In order to be acquainted with GPU acceleration relatively simple interaction, Lennard­Jones potential model implemented with force smoothening applied in LAMMPS GPU package. The reference data to validate GPU algorithms is derived from the CPU version. The method of checking the numerical accuracy of molecular dynamics trajectories based on exponential divergence analysis has been implemented and applied.