Abstract
Nonsmooth optimal control problems (OCPs) constrained by partial differential equations (PDEs) present significant computational challenges due to the non-differentiability of objective functions and the high cost of repeated PDE evaluations. This paper presents two synergistic strategies to accelerate the solution of such problems: a theoretical refinement of the Primal-Dual Hybrid Gradient (PDHG) algorithm that permits significantly larger step sizes, and the integration of DeepONets as neural operator surrogates. We demonstrate that these methods collectively reduce iteration counts and computational overhead while maintaining rigorous convergence and mesh independence.
Introduction
Optimal control problems (OCPs) are fundamental to science and engineering, aiming to identify a control input $u$ that minimizes a cost functional while satisfying a PDE constraint $y = Su$. In many practical applications, these problems are nonsmooth due to box constraints on the control or sparsity-promoting regularizers (e.g., $L^1$-norms).
Solving these problems typically involves two major bottlenecks. First, the nonsmoothness prevents the use of standard gradient-based methods, necessitating proximal algorithms like the Primal-Dual Hybrid Gradient (PDHG) method. Second, each iteration of such algorithms requires solving the forward PDE and its adjoint, which is computationally prohibitive for complex or time-dependent systems. This work addresses both bottlenecks through algorithmic optimization and operator learning.
Methodology
1. Theoretical Acceleration: Enlarged Step Sizes
The PDHG algorithm, also known as the Chambolle-Pock algorithm, is a standard tool for nonsmooth optimization. Its convergence traditionally relies on a restrictive step-size condition: the product of the primal step $r$ and dual step $s$ must satisfy $rs < 1/\|S\|^2$, where $\|S\|$ is the operator norm of the PDE solver.
The authors exploit the strong convexity often present in OCPs (typically arising from the $L^2$ regularization of the control) to prove a more permissive convergence condition. They demonstrate that the step sizes can be enlarged to:
$$r \cdot s < \frac{4 + 2\alpha r}{3} \cdot \frac{1}{\|S\|^2}$$
where $\alpha$ represents the strong convexity parameter. This "Improved" PDHG (PD-I) allows for larger, more aggressive updates, leading to faster convergence compared to the "Classic" PDHG (PD-C).
2. Operator Learning with DeepONets
To address the cost of the PDE solver itself, the authors replace traditional Finite Element Method (FEM) solvers with DeepONets—a class of neural operators designed to learn mappings between infinite-dimensional function spaces.
The framework involves training two separate neural networks:
- The Forward Operator ($S$): Maps the control $u$ to the state $y$.
- The Adjoint Operator ($S^*$): Maps the dual variables to the adjoint state.
Unlike standard neural networks that learn a mapping for a specific discretization, DeepONets learn the underlying operator. This makes the optimization "mesh-free" and allows the pre-trained model to be reused for different target states $y_d$ without retraining. Once trained, the "solver" becomes a simple forward pass through the network, providing near-instantaneous evaluations.
Experimental Setup and Data Generation
The methodology was evaluated on both elliptic (steady-state) and parabolic (time-dependent) control problems. Since standard datasets for these specific OCPs are unavailable, the authors generated synthetic training data:
- Input Functions: Controls were sampled from Gaussian Random Fields (GRF) using a Riesz kernel to ensure a diverse and representative training set.
- Ground Truth: High-fidelity traditional solvers (Fourier-based or FEM) provided the "exact" solutions for the $(u, y)$ pairs.
- Benchmarks: The performance of the DeepONet-based PDHG was compared against traditional FEM-based PDHG in terms of accuracy, iteration count, and robustness.
Key Results and Discussion
The integration of enlarged step sizes and operator learning yielded significant performance gains:
- Convergence Speed: In parabolic control benchmarks, the PD-I variant (using enlarged step sizes) converged approximately three times faster than the classic PD-C variant.
- Approximation Accuracy: The DeepONet surrogates achieved relative errors between $10^{-2}$ and $10^{-3}$. While this introduces a small approximation gap compared to FEM, the resulting control solutions remain visually and functionally indistinguishable from the ground truth.
- Robustness and Mesh Independence: The methods demonstrated consistent performance across various mesh densities and regularization parameters ($\alpha$ and $\mu$). This suggests that the neural operator successfully captures the continuous physics of the problem rather than just interpolating on a grid.
Conclusion and Future Directions
This work demonstrates that combining rigorous optimization theory with modern deep learning can significantly alleviate the computational burden of nonsmooth optimal control. The enlarged step-size condition provides a "free" speedup for strongly convex problems, while DeepONets offer a path toward real-time control by replacing expensive PDE solvers.
Future research will focus on extending this framework to nonlinear PDE constraints, such as the Navier-Stokes equations, where the computational savings of neural operators would be even more impactful. Additionally, exploring hybrid approaches—where neural operators provide a fast initial guess followed by FEM refinement—could bridge the gap between high-speed approximation and high-precision requirements.
Summary updated