GDGen: A gradient descent-based methodology for the generation of optimized spatial configurations of customized clusters in computational simulations

Abstract

In this study, Gradient Descent Generation (GDGen) is presented, an innovative methodological framework that utilizes gradient descent algorithms to create dense, non-overlapping configurations of multiple, user-customized clusters and shapes. This technique is crucial for the accuracy and efficacy of molecular dynamics (MD) simulations, finite element analyses, and a multitude of scientific applications where precise spatial arrangement is paramount. GDGen intricately minimizes a loss function tailored to assess spatial overlaps and guide the arrangement process.

The implementation of GDGen is encapsulated in Pygdgen, a Python package developed to generate intricate atomic configurations, particularly excelling in scenarios involving dense clustering and unconventional geometries. Pygdgen ensures efficient arrangement processes through its optimized coding structure and GPU acceleration capabilities. Its adaptability is evidenced by its application in various fields, from material science and chemistry to urban planning and mechanical design, for arranging complex structures within constrained spaces.