Improving structure search with hyperspatial optimization and TETRIS seeding.

Advanced structure prediction methods developed over the past decades include an unorthodox strategy of allowing atoms to displace into extra dimensions. A recently implemented global optimization of structures from hyperspace (GOSH) has shown promise in accelerating the identification of global minima on potential energy surfaces defined by simple interatomic models. In this study, we extend the GOSH formalism to more accurate Behler-Parrinello neural network (NN) potentials, make it compatible with efficient local minimization algorithms, and test its performance on nanoparticles and crystalline solids. For clusters modeled with NN potentials, four-dimensional optimization offers fairly modest improvement in navigating geometric relaxation pathways and incurs increased computational cost largely offsetting the benefit, but it provides a significant advantage in facilitating atom swaps in nanoalloys. In comparison, the introduction of a moderate, controlled bias for generating more physically sensible starting configurations, achieved via TETRIS-inspired packing of atomic blocks, has a more direct impact on the efficiency of global structure searches. The benchmarked systems are Lennard-Jones clusters, Au or Cu-Pd-Ag nanoparticles and binary Sn alloys described by NN potentials, and compounds with covalent B or BC frameworks modeled with density functional theory.