Adaptive multigrid strategy for geometry optimization of large-scale three dimensional molecular mechanics

Lei Zhang, Kejie Fu, Mingjie Liao, Yangshuai Wang, Jianjun Chen, Lei Zhang 0027Journal of Computational Physics 2023 · DOI

TL;DR

This paper introduces a novel adaptive multigrid strategy for optimizing the geometry of large-scale molecular mechanics systems. By combining adaptive mesh refinement with a blended ghost force correction (BGFC) coarse-grid approximation, the method achieves significant speedups (up to five-fold) compared to brute-force optimization, enabling the efficient simulation of systems with millions of atoms.

What Problem Does This Paper Address?

Geometry optimization is a fundamental and computationally intensive task in materials science and molecular mechanics, aiming to find the lowest energy configuration of atoms. For large systems, especially those with complex defects like dislocations and cracks, traditional "brute-force" optimization methods become prohibitively slow due to their computational complexity scaling linearly or even super-linearly with the number of atoms. Existing methods often struggle to balance accuracy and efficiency, particularly in three dimensions. This paper addresses the need for a scalable and efficient geometry optimization method that can handle large-scale, complex molecular systems.

Methods

The core of the proposed method is an adaptive multigrid strategy. This involves:

  • Multigrid Framework: Leveraging a hierarchical approach where computations are performed on multiple levels of mesh resolution.
  • Coarse-Grid Approximation: Instead of performing brute-force calculations on all atoms at each level, the method uses a Blended Ghost Force Correction (BGFC) approximation. This is an atomistic/continuum coupling technique that significantly reduces the number of degrees of freedom on coarser levels.
  • Adaptive Mesh Refinement: The mesh is not fixed but dynamically refined in regions where it's most needed, guided by a posteriori error estimators. This ensures that computational resources are focused on areas with significant atomic displacement or stress.
  • 3D Mesh Generation: The paper utilizes advanced 3D mesh generation techniques that adapt to the underlying crystalline structure, which is crucial for accurate representation of defects and interfaces in three dimensions.
  • Oneway Multigrid: The strategy employs a "oneway" multigrid approach, meaning information flows primarily from coarser to finer levels, simplifying the implementation and comparison with existing methods.

Datasets

Not discussed in the paper. The paper focuses on simulating crystalline defects in tungsten (W) using established interatomic potentials (EAM).

Key Insights and Ideas

  • BGFC as an Efficient Coarse-Grid Solver: The BGFC method, when used as a coarse-grid approximation in a multigrid scheme, provides a good balance between accuracy and computational cost, enabling sublinear scaling for geometry optimization.
  • Adaptivity is Crucial for Complex Defects: For defects with longer-range elastic fields (like dislocations), adaptive mesh refinement is essential to maintain efficiency and accuracy, preventing the method from degrading to near-linear scaling.
  • 3D Mesh Generation is a Key Enabler: The development of robust 3D mesh generation and adaptation techniques tailored for crystalline structures is critical for extending these multiscale methods to three-dimensional problems.
  • Modeling Error Dominates: For coarse-grid problems, the accuracy of the approximation (modeling error) can become the limiting factor, suggesting that excessively high accuracy in coarse-grid solutions might not always translate to significant overall speedup.

Key Results

  • Five-fold Acceleration: For systems with up to 100 million atoms, the proposed OAM-BGFC strategy achieves a five-fold acceleration in CPU time compared to brute-force optimization.
  • Sublinear Scaling: The method demonstrates sublinear scaling (e.g., O(N^0.78) for single vacancy) with respect to the number of atoms (N), a significant improvement over the linear or super-linear scaling of brute-force methods.
  • Efficiency for Complex Defects: While point defects (like vacancies) show substantial speedups, the method also provides significant benefits for more complex defects like micro-cracks and edge dislocations, reducing complexity rates considerably. For edge dislocations, savings of around 60% are observed for systems with ten million atoms.
  • Mesh Generation Cost is Negligible: The computational cost of mesh generation and adaptation is found to be less than 1% of the total CPU time, highlighting the efficiency of their implemented generator.

Limitations and Future Work

  • Rigorous a Priori Error Estimates for Dislocations: The paper acknowledges that rigorous a priori error estimates for BGFC approximation for general straight dislocations are still lacking. Future work will involve studying equivalent ghost force removal formulations and developing these estimates.
  • More Complex Multigrid Strategies: The current work focuses on a "oneway" multigrid strategy. Future research will explore more complex multigrid schemes like Full Multigrid (FMG) and Full Approximation Scheme (FAS) to potentially achieve even greater efficiency.
  • Adaptive Strategies for Complex Defects: Extending the adaptive multigrid strategies to more complex crystalline defects such as cracks, partial dislocations, and grain boundaries remains a significant challenge, particularly in constructing and implementing the corresponding a/c coupling approximations.
  • Choice of Predictor in BGFC: The paper uses a trivial predictor (û₀ = 0) for the BGFC method. Investigating non-trivial choices for this predictor, especially for cracks and dislocations, could lead to improved convergence rates and sublinear scaling.

Summary updated

BibTeX
@article{zhang2023adaptive,
  author = {Lei Zhang and Kejie Fu and Mingjie Liao and Yangshuai Wang and Jianjun Chen and Lei Zhang},
  title = {{Adaptive multigrid strategy for geometry optimization of large-scale three dimensional molecular mechanics}},
  journal = {Journal of Computational Physics},
  year = {2023},
  doi = {10.1016/j.jcp.2023.112113},
}
Figure