Improving structure search with hyperspatial optimization and TETRIS seeding.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Advanced structure prediction methods are crucial for materials discovery.
• Global optimization of structures from hyperspace (GOSH) offers a novel approach.
• Extending GOSH to accurate potentials like Behler-Parrinello neural networks (NN) is a key challenge.

Purpose of the Study:

• To extend the GOSH formalism to Behler-Parrinello neural network potentials.
• To assess the performance of GOSH with NN potentials on nanoparticles and crystalline solids.
• To compare the efficacy of 4D optimization versus biased initial configuration generation.

Main Methods:

• Implementation of GOSH with Behler-Parrinello NN potentials.
• Integration with efficient local minimization algorithms.
• Benchmarking on Lennard-Jones clusters, Au/Cu-Pd-Ag nanoparticles, binary Sn alloys, and B/BC frameworks using NN potentials and density functional theory.

Main Results:

• Four-dimensional optimization provided modest improvements for cluster relaxation but increased computational cost.
• GOSH with NN potentials significantly aided atom swaps in nanoalloys.
• TETRIS-inspired packing for generating initial configurations showed a more direct impact on global structure search efficiency.

Conclusions:

• Extending GOSH to NN potentials is feasible but requires careful consideration of computational cost.
• Biased initial configuration generation strategies offer a more effective route to accelerate global structure searches.
• The study provides insights into optimizing structure prediction for complex materials systems.