High order spatial discretization for variational time implicit schemes: Wasserstein gradient flows and reaction-diffusion systems

Guosheng Fu, Stanley J. Osher, Wuchen Li

Introduction

Many physical, biological, and chemical processes are characterized by dissipative dynamics, where a system evolves toward equilibrium by minimizing a specific energy functional. Mathematically, these processes are often modeled as gradient flows. In the context of optimal transport, the Jordan-Kinderlehrer-Otto (JKO) scheme provides a rigorous variational framework for these flows, particularly those governed by the Wasserstein distance. However, implementing the JKO scheme with high-order spatial accuracy and computational efficiency has historically been a significant challenge.

This paper introduces a robust numerical framework that addresses these hurdles. By combining high-order finite element methods (FEM) with a variational optimization approach, the authors provide a stable and accurate methodology for simulating complex dissipative systems, including nonlinear Fokker-Planck equations and multi-species reaction-diffusion systems like the Gray-Scott model.

Methodology

The proposed framework treats the evolution of partial differential equations (PDEs) as a sequence of optimization problems. The approach is built upon three primary pillars:

1. One-Step Relaxation

To avoid the computational complexity of solving the full JKO distance problem at every step, the authors utilize a "one-step relaxation." This simplification maintains first-order temporal accuracy while significantly reducing the optimization burden, making the variational approach more tractable for complex geometries and high-dimensional systems.

2. The ALG2 Algorithm

The core of the optimization process is the Augmented Lagrangian method (ALG2). This "splitting" technique decomposes the complex nonlinear problem into two manageable steps:

  • Step A: A linear reaction-diffusion solve, which is standard and computationally efficient.
  • Step B/C: A nonlinear update that can be computed point-wise at every quadrature point. This point-wise efficiency is a critical insight, as it bypasses the need for expensive global nonlinear solvers and allows for massive parallelization.

3. High-Order Finite Elements (FEM)

The spatial discretization employs high-order $H^1$-conforming elements for potential variables and $L^2$-conforming discontinuous polynomials for density and flux. The researchers found that using equal-order polynomials for all variables (potential, density, and flux) yielded superior accuracy compared to traditional mixed-order approaches. This allows the simulation to achieve $(k+1)$-th order convergence in space for a polynomial degree $k$.

Numerical Experiments and Benchmarks

The framework was validated against several classic and challenging PDE benchmarks to test its versatility and robustness:

  • Nonlinear Fokker-Planck Equation: Used primarily to verify the high-order spatial convergence rates.
  • Aggregation-Drift-Diffusion Equations: Demonstrated the method's capability in handling nonlocal interaction kernels, often found in swarming and biological modeling.
  • Fisher-KPP Equation: A benchmark for population dynamics and flame propagation, testing the method's performance in non-convex energy landscapes.
  • Reversible Reaction-Diffusion Systems: The authors successfully simulated the 4-species Gray-Scott model. Notably, they were able to capture intricate "spot" and "ring" patterns that previous numerical attempts had missed, attributing earlier failures to inappropriate parameter choices rather than mathematical limitations of the gradient flow formulation.

Key Insights

Unconditional Stability

For convex problems—where the energy functional is convex and the mobility is concave—the method is unconditionally stable. This allows researchers to utilize large time steps (e.g., $\Delta t = 1$) without risking numerical divergence, which is particularly useful for reaching steady-state solutions quickly.

Efficiency through Point-wise Updates

The realization that the nonlinear optimization step in the ALG2 algorithm can be reduced to point-wise calculations is a major contribution. This ensures that the high-order spatial accuracy does not come at the cost of prohibitive computational overhead.

Entropy Dissipation

The numerical results consistently demonstrate the monotone decay of energy, or entropy dissipation, as predicted by the underlying theory of gradient flows. This thermodynamic consistency is a hallmark of the variational approach.

Limitations and Future Directions

While the framework is highly accurate in space, it is currently limited to first-order accuracy in time. In scenarios involving fast transients, this may necessitate smaller time steps to maintain temporal precision. Additionally, the guarantee of unconditional stability is lost for non-convex problems, such as the Fisher-KPP equation, requiring more careful parameter tuning.

Future work aims to extend this variational framework to higher-order temporal discretizations. Furthermore, the authors intend to apply these high-order spatial methods to more complex fluid dynamics and regularized conservation laws, expanding the reach of Wasserstein gradient flow simulations in the physical sciences.

Summary updated

BibTeX
@article{fu2023high,
  author = {Guosheng Fu and Stanley J. Osher and Wuchen Li},
  title = {{High order spatial discretization for variational time implicit schemes: Wasserstein gradient flows and reaction-diffusion systems}},
  journal = {Journal of Computational Physics},
  year = {2023},
  doi = {10.1016/j.jcp.2023.112375},
}
Figure