Convex splitting method for the calculation of transition states of energy functional

Xiang Zhou 周翔, Shuting Gu, Xiang Zhou

TL;DR

This paper introduces a novel numerical method, the convex splitting method, to efficiently find "transition states" (index-1 saddle points) of energy functionals. This method significantly improves computational efficiency by allowing for larger time steps, which is crucial for simulating complex physical and chemical systems.

What Problem Does This Paper Address?

Many physical phenomena, from phase transitions in materials to chemical reactions, are governed by energy landscapes. Understanding these landscapes requires identifying not only stable configurations (local minima) but also the pathways between them. These pathways are often characterized by "transition states," which are unstable critical points (index-1 saddle points) on the energy surface. Finding these saddle points is computationally challenging, especially for complex, spatially extended systems described by partial differential equations. Existing methods often struggle with stability and require very small time steps, making simulations prohibitively slow. This paper aims to develop a more efficient and stable numerical approach for locating these critical transition states.

Methods

The core of this paper is the adaptation of the "convex splitting method" for finding saddle points. The convex splitting method, originally designed for stable states, works by splitting an energy functional into convex and concave parts, allowing for unconditionally stable numerical schemes with large time steps.

The authors extend this idea to the "iterative minimization formulation" (IMF), a standard framework for finding saddle points. The IMF breaks down the saddle point problem into a series of minimization subproblems. The key innovation here is applying the convex splitting technique to these subproblems.

Specifically, they:

  • Derive convex splitting forms for auxiliary functionals: They show how to construct convex splitting forms for the auxiliary functionals used within the IMF, automatically adapting existing convex splitting techniques.
  • Ensure time-explicit discretization: A crucial aspect is ensuring that the numerical schemes remain linear and time-explicit, avoiding the need to solve complex nonlinear equations at each step. This is achieved by carefully selecting different convex splitting forms for different terms within the auxiliary functional.
  • Apply to gradient flows: The method is implemented by approximating the steady state of a gradient flow associated with the auxiliary functional.

The authors then apply this method to two key examples:

  1. 1D Ginzburg-Landau energy functional: This model is relevant to phase transitions and is studied under both L² (Allen-Cahn type) and H⁻¹ (Cahn-Hilliard type) metrics.
  2. 2D Landau-Brazovskii energy functional: This model describes phase transitions in diblock copolymers, studied in the H⁻¹ metric.

Datasets

Not discussed in the paper. The paper focuses on numerical methods and their application to established theoretical models.

Key Insights and Ideas

  • Convex splitting is adaptable: The paper demonstrates that the benefits of convex splitting (stability, large time steps) can be successfully extended from finding stable states to finding unstable saddle points within the IMF framework.
  • Efficiency gain is significant: By allowing larger time steps, the new method dramatically reduces the total computational cost for locating transition states, especially for complex systems.
  • Avoiding nonlinear solvers is key: The authors emphasize the importance of constructing schemes that avoid solving nonlinear systems at each time step, which is a major bottleneck in many numerical methods. Their approach achieves this by cleverly combining different convex splitting forms.
  • Min-mode selection is critical: The paper highlights a subtle but important issue in selecting the correct "min-mode" for saddle point search, particularly when dealing with degenerate cases (e.g., due to symmetries), and proposes a practical solution.

Key Results

  • Unconditional energy stability: The proposed convex splitting schemes for saddle point search are shown to be unconditionally energy stable, meaning they remain stable for any time step size.
  • Quadratic convergence of IMF: The underlying IMF framework is validated to exhibit quadratic convergence for finding saddle points, as expected.
  • Significant computational speedup: Numerical experiments show that the convex splitting method (CS) significantly outperforms the non-convex splitting (nCS) method in terms of total computational cost, especially for larger time step sizes. For instance, in the 1D Ginzburg-Landau case, the CS scheme shows a much better stability and efficiency profile compared to nCS when large time steps are used.
  • Robustness: The method is tested on various initial conditions and boundary conditions, demonstrating its robustness.
  • Application to complex systems: The method is successfully applied to both 1D and 2D models relevant to materials science and phase transitions, showcasing its practical utility.

Limitations and Future Work

  • Local convexity: While the method relies on convex splitting, the authors note that global convexity of all terms might not always be possible. They rely on local convexity within the relevant solution space, which has no strict theoretical foundation but works in practice for the tested examples.
  • Min-mode selection: The paper discusses a potential issue with min-mode selection in degenerate cases and offers a practical solution. However, further theoretical investigation into this aspect might be beneficial.
  • Stability of new schemes: The authors suggest that extending their approach to other stable numerical methods for gradient dynamics could be a future research direction, but acknowledge that the theoretical stability of these new schemes for L would need to be investigated.

Summary updated

Abstract

Among numerical methods for partial differential equations arising from steepest descent dynamics of energy functionals (e.g., Allen-Cahn and Cahn-Hilliard equations), the convex splitting method is well-known to maintain unconditional energy stability for a large time step size. In this work, we show how to use the convex splitting idea to find transition states, i.e., index-1 saddle points of the same energy functionals. Based on the iterative minimization formulation (IMF) for saddle points (SIAM J. Numer. Anal., vol. 53, p1786, 2015), we introduce the convex splitting method to minimize the auxiliary functional at each cycle of the IMF. We present a general principle of constructing convex splitting forms for these auxiliary functionals and show how to avoid solving nonlinear equations. The new numerical scheme based on the convex splitting method allows for large time step sizes. The new methods are tested for the one dimensional Ginzburg-Landau energy functional in the search of the Allen-Cahn or Cahn-Hilliard types of transition states. We provide the numerical results of transition states for the two dimensional Landau-Brazovskii energy functional for diblock copolymers.
BibTeX
@article{2018convex,
  author = {Xiang ZHOU 周翔 and Shuting Gu and Xiang Zhou},
  title = {{Convex splitting method for the calculation of transition states of energy functional}},
  journal = {Journal of Computational Physics},
  year = {2018},
  doi = {10.1016/j.jcp.2017.10.028},
}
Figure