Introduction
Identifying discontinuous diffusion coefficients in elliptic equations is a fundamental yet challenging inverse problem in partial differential equations (PDEs). This task is critical for applications such as mapping underground rock permeability or determining the heat conductivity of composite materials. However, the problem is inherently ill-posed and nonlinear, requiring specialized techniques to preserve sharp boundaries between different material zones. While Total Variation (TV) regularization is a standard tool for maintaining these discontinuities, its non-smooth nature often forces researchers to use smoothing approximations that can compromise reconstruction quality.
The ADMM-Newton-CNN Framework
To address these challenges without sacrificing mathematical integrity, this paper proposes a hybrid numerical strategy: the ADMM-Newton-CNN approach. This framework combines the Alternating Direction Method of Multipliers (ADMM) with a second-order Newton method and a deep Convolutional Neural Network (CNN).
1. ADMM Splitting
The core of the strategy is using ADMM to decompose the complex optimization problem into two simpler subproblems. This splitting allows the TV regularization term to remain in its original, non-smooth form, effectively exploiting the "hidden" convexity of the model when gradient observations are available.
2. Physics-Based Subproblem: Active-set Newton Method
The first subproblem handles the underlying physics of the elliptic equation. The authors employ a second-order Newton method to solve this efficiently. To manage large-scale discretizations, they utilize a Schur complement reduction combined with a "lumped mass matrix." This approach significantly reduces the computational complexity of the linear systems involved.
The gradient computation, as detailed in the subroutine above, is a critical step in the Newton iteration, involving the solution of the state equation and the subsequent evaluation of the gradient components.
3. Regularization Subproblem: Deep CNN
The second subproblem is essentially a denoising task. Instead of traditional iterative solvers, the framework integrates a pre-trained Deep Convolutional Neural Network (DnCNN). In this context, the CNN acts as a proximal operator, efficiently cleaning the reconstruction while preserving the sharp edges characteristic of the TV model.
Experimental Results
The method was evaluated using synthetic benchmarks with varying levels of complexity and noise:
- Geometries: Tests included simple straight-line discontinuities and complex domains with circular and right-angled features.
- Noise Robustness: The approach remained stable under Gaussian noise levels up to 10% ($\delta = 0.1$).
- Scalability: On a $1024 \times 1024$ grid (over a million unknowns), the algorithm converged in approximately 10-15 minutes on a standard desktop. The number of Newton iterations per ADMM step remained consistently low, demonstrating excellent mesh-independence.
Conclusion and Future Directions
The ADMM-Newton-CNN approach demonstrates that deep learning can be rigorously integrated into traditional optimization frameworks as a mathematical component rather than a "black box." By avoiding the smoothing of the TV term and utilizing efficient physics solvers, the method achieves high accuracy and speed.
Future work will focus on extending this framework to the non-convex case where only function values (rather than gradients) are observed, as well as adapting the CNN components for irregular 3D domains.
Summary updated