Abstract
A state-of-the-art deep domain decomposition method (D3M) based on the variational principle is proposed for partial differential equations (PDEs). The solution of PDEs can be formulated as the solution of a constrained optimization problem, and we design a multi-fidelity neural network framework to solve this optimization problem. Our contribution is to develop a systematical computational procedure for the underlying problem in parallel with domain decomposition. Our analysis shows that the D3M approximation solution converges to the exact solution of underlying PDEs. Our proposed framework establishes a foundation to use variational deep learning in large-scale engineering problems and designs. We present a general mathematical framework of D3M, validate its accuracy and demonstrate its efficiency with numerical experiments.
Introduction: The Challenge of Large-Scale PDEs
Partial Differential Equations (PDEs) are the backbone of physics and engineering, but solving them for large-scale, complex systems remains computationally grueling. Traditional methods like Finite Element Analysis (FEM) are robust but can become prohibitively expensive as domain complexity increases.
Recently, "physics-informed" deep learning has emerged as a mesh-free alternative. However, global neural networks face significant hurdles: they often struggle with the "curse of dimensionality," find it difficult to satisfy complex boundary conditions, and are prone to overfitting when a single network attempts to learn a solution across a massive, intricate domain. The Deep Domain Decomposition Method (D3M) bridges the gap between classical numerical analysis and modern AI to create a solver that is both scalable and accurate.
The D3M Methodology
D3M rests on three main pillars designed to enable parallel computation and improve convergence:
1. Domain Decomposition
Following the classical Schwarz alternating method, the global domain is split into overlapping subdomains. Each subdomain is assigned its own neural network, allowing the model to manage complexity at the local level. This "divide and conquer" strategy reduces the hypothesis space for each network, acting as a form of regularization that significantly reduces the risk of overfitting compared to a single massive network.
2. Variational Principle and Multi-Fidelity Networks
Instead of simply minimizing the residual of the PDE (as in standard Physics-Informed Neural Networks), D3M formulates the PDE as a constrained optimization problem. It utilizes a "mixed residual loss" based on the Hellinger-Reissner principle, solving for both the solution ($u$) and the flux ($\tau$) simultaneously. This multi-fidelity approach ensures that the physical constraints are embedded directly into the learning process.
3. Boundary Function Transformation
To handle inhomogeneous boundary conditions, which often trip up standard neural networks, the authors introduce a smooth boundary function $g(x,y)$. This transforms the problem into one with homogeneous boundary conditions, making it much easier for the local networks to converge during training.
Experimental Evaluation and Results
The paper evaluates D3M on fundamental mathematical benchmarks, comparing it against the Deep Ritz Method (DRM) and traditional Finite Element Methods (FEM):
- Poisson’s Equation: In a 2D square domain test, D3M achieved a relative error of 0.0045, compared to 0.0157 for the standard DRM using the same number of neurons and residual blocks.
- Schrödinger Equation: For a 2D infinite potential well, D3M showed a marked improvement in capturing probability density, with a relative error of 0.0091 versus DRM's 0.0334.
One of the most significant findings is that D3M performs better at capturing peak values in the solution (such as the center of a wave function) compared to global methods, which tend to "smooth out" these critical features.
Conclusion and Future Directions
D3M establishes a foundation for using variational deep learning in large-scale engineering. By design, the method is highly parallelizable; the "offline" training phase for local networks requires no communication between subdomains, making it perfectly suited for distributed computing.
While powerful, current limitations include sensitivity to weight initialization and the empirical selection of Lagrange multipliers in the loss function. Future research will focus on extending D3M to time-dependent PDEs and refining the communication steps between subdomains to further accelerate global convergence.
Summary updated