Numerical Methods for the Optimal Control of Scalar Conservation Laws

Michaël Herty, Lorenzo Pareschi, Sonja SteffensenarXiv 2012 · DOI · arXiv

Abstract

We are interested in a class of numerical schemes for the optimization of nonlinear hyperbolic partial differential equations. We present continuous and discretized relaxation schemes for scalar, one-dimensional conservation laws. We present numerical results on tracking type problems with nonsmooth desired states and convergence results for higher-order spatial and temporal discretization schemes.

Introduction

Optimal control problems governed by scalar conservation laws, such as the Burgers’ equation, present significant numerical challenges. The inherent nonlinearity of these equations often leads to the formation of shocks and wave interactions. From a mathematical perspective, the objective function to be minimized is frequently non-differentiable in the presence of such discontinuities. Furthermore, standard numerical solvers often struggle with the "stiffness" introduced by these nonlinearities, making it difficult to obtain accurate gradients (adjoints) for optimization. This work provides a systematic framework for discretizing these problems to ensure high-order accuracy for both the state and adjoint equations, even when the underlying physics is complex.

Methodology

The proposed approach rests on two primary technical pillars: relaxation methods and Implicit-Explicit (IMEX) Runge-Kutta discretization.

Relaxation Method

Instead of directly solving the nonlinear conservation law, the authors utilize the Jin-Xin relaxation model. This model replaces the nonlinear PDE with a system of linear transport equations coupled with a stiff source term. As the relaxation parameter $\epsilon$ approaches zero, the solution of this linear system recovers the original nonlinear physics. This transformation significantly simplifies the numerical handling of the problem.

IMEX Runge-Kutta Discretization

To solve the resulting relaxation system, Implicit-Explicit (IMEX) schemes are employed. The explicit component handles the linear transport (wave movement), while the implicit component manages the stiff source term (relaxation back to nonlinear equilibrium).

Discretize-then-Optimize

The authors follow a "discretize-then-optimize" strategy, deriving the discrete adjoint equations directly from the discretized forward system. A rigorous analysis establishes the conditions for Runge-Kutta coefficients to ensure that the entire optimality system achieves first, second, or third-order accuracy.

Experimental Setup

The framework is validated using a standard numerical benchmark: the Burgers' equation. The specific task is a tracking problem where the goal is to identify an initial state $u_0$ that evolves to match a target state $u_d$ at time $T$. The spatial interval is $[0, 2\pi]$, with a target state derived from $u_d(0, x) = 0.5 + \sin(x)$.

Key Results and Discussion

The numerical experiments demonstrate several key strengths of the IMEX relaxation approach.

Tracking Performance

The method successfully converges to the target state, even when the solution contains nonsmooth features like shocks. Figure 1 illustrates the alignment between the optimal solution and the target density, starting from a uniform initial state.

Comparison of the initial, optimal, and target density profiles for the Burgers' equation tracking problem.

Grid Independence and Efficiency

A significant finding is the "grid independence" of the optimization process. As shown in Figure 2, the number of iterations required for convergence remains remarkably stable (between 41 and 44) as the spatial grid is refined from $N=100$ to $N=300$. This indicates that the optimization landscape is well-captured by the discretization. Additionally, CPU times scale predictably with the number of grid points, confirming the efficiency of the IMEX approach.

Performance metrics showing the number of iterations and CPU time for different grid sizes (N).

High-Order Accuracy

The authors successfully defined order conditions for up to third-order temporal accuracy for the optimality system. This ensures that the computed gradients are highly accurate, facilitating efficient optimization.

Conclusion and Future Work

This paper introduces a robust framework for the optimal control of scalar conservation laws using high-order IMEX relaxation schemes. While the current analysis focuses on scalar laws and often assumes smooth data for formal proofs, the numerical results demonstrate effectiveness even with shocks.

Future research will focus on extending these methods to systems of conservation laws, such as the Euler equations of gas dynamics. Additionally, investigating properties like "asymptotic preservation" and strong stability for these specific optimal control discretizations remains a priority.

Summary updated

Abstract

We are interested in a class of numerical schemes for the optimization of nonlinear hyperbolic partial differential equations. We present continuous and discretized relaxation schemes for scalar, one-- conservation laws. We present numerical results on tracking type problems with nonsmooth desired states and convergence results for higher--order spatial and temporal discretization schemes.
BibTeX
@misc{herty2012numerical,
  author = {Michaël Herty and Lorenzo Pareschi and Sonja Steffensen},
  title = {{Numerical Methods for the Optimal Control of Scalar Conservation Laws}},
  year = {2012},
  eprint = {1207.3671},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1207.3671},
}
Figure