Source-linked AI summary
D3M: A deep domain decomposition method for partial differential equations
Ke Li, Kejun Tang, Tianfan Wu, Qifeng Liao
TL;DR
Large-scale engineering PDE problems can be costly to solve, while accuracy and generalization are difficult to balance. D3M applies variational deep learning with domain decomposition and parallel subdomain procedures. The authors report high-accuracy solutions for general parabolic PDEs.
Problem
Practical engineering PDE problems can be expensive to solve, and balancing accuracy with generalization is difficult.
Source: §1 · p. 1Method
D3M trains local neural networks under a variational formulation on decomposed subdomains and performs their procedures in parallel.
Source: §1 · pp. 1–2Results
The authors demonstrate that D3M can solve general parabolic PDEs with high accuracy.
Source: §7 · p. 14Takeaways & Limitations
The authors present parallel computation as a foundation for physics-constrained deep learning in large-scale engineering simulations or designs.
Source: §7 · p. 14In practice, D3M faces bottlenecks from neural-network initialization, interface-approximation choices, and selecting Lagrangian multipliers without prior guidance.
Source: §7 · p. 14
Abstract
from arXiv · showhide
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.
1. Introduction.
Deep-learning approaches address PDE solving, while practical engineering problems motivate domain decomposition to manage complexity and expensive hardware demands. The section introduces D3M, a variational deep-learning solver that computes in parallel across physical subdomains.
1. Introduction.
- Prior PDE deep-learning approaches include variational, physics-informed, model-discovery, mesh-free, and Bayesian methods.The introduction names DRM, physics-informed neural networks, PDE-Net, DGM, and Bayesian deep convolutional encoder-decoder networks.Source: §1 · p. 1
- Domain decomposition is presented as a way to manage complex engineering systems whose solutions can require expensive hardware and difficult accuracy–generalization trade-offs.Source: §1 · p. 1
- D3M combines variational deep learning with domain decomposition to parallelize computation across physical subdomains.The paper describes independent local model training followed by assembling the global solution from precomputed local information.Source: §1 · pp. 1–2
- Splitting the physical domain lets each network focus locally, which the authors say can reduce overfitting risk.Source: §1 · p. 1
- D3M trains networks under a variational formulation and does not require given data, with potential use in complex, high-dimensional problems.Source: §1 · p. 2
- D3M performs procedures for different subdomains in parallel, a feature the authors say helps with large-scale and multidisciplinary problems.Source: §1 · p. 2
2. Overlapping domain decomposition.
The section introduces overlapping domain decomposition through the classical Schwarz alternating method, with a two-subdomain partition illustrated and formalized.
2. Overlapping domain decomposition.
- The Schwarz method is identified as a classical domain decomposition approach for PDEs, with later variants improving it.Source: §2 · p. 2
- The section takes a classical Poisson equation as its starting problem.Source: §2 · p. 2
- The illustration partitions the domain into two overlapping subdomains; the text denotes their shared region and the two interfaces.The passage defines Γ1 and Γ2 as the respective subdomain interfaces and Ω1,2 as the overlap.Source: §2 · p. 2, Figure 2.1 · p. 2
- The Schwarz alternating method starts from an initial guess and defines local solution sequences through an iteration algorithm.Source: §2 · p. 2
3. Deep domain decomposition method with variational principle.
The section turns the Poisson problem into a variational optimization and uses neural networks to approximate its solution and flux. D3M applies this approach across overlapping subdomains, whose local solutions are assembled into a global solution.
3.1. Variational principle.
- The Poisson equation is formulated as an optimization problem, with a mixed residual formulation introducing flux as an additional variable.The Lagrangian uses q as a Lagrange multiplier.Source: §3.1 · p. 3
3.2. Variational principle informed neural networks.
- Neural networks approximate the solution u and flux τ, converting the infinite-dimensional minimization into a trainable finite-dimensional optimization.In implementation, one network can provide both outputs.Source: §3.2 · p. 3
3.3. Implementation details for neural networks.
- The network takes x and y as inputs and outputs u and τ; the implementation uses tanh activation, a ResNet architecture, and L-BFGS optimization.The stated learning rate is 0.5.Source: §3.3 · p. 4, Figure 3.1 · p. 4
3.4. Deep domain decomposition method.
- A boundary function matches u on the boundary and decays sharply away from it, while the method also uses improved junction sampling for inhomogeneous boundary conditions.The text says domain decomposition reduces Monte Carlo sampling variance for the same sample count because each sampled area is smaller.Source: §3.4 · pp. 4–5
- D3M divides the domain into overlapping subdomains, trains local neural-network solutions through the variational principle, and assembles them into a global solution.The algorithm runs subdomain procedures in parallel and iterates while its accuracy threshold is unmet.Source: §3.4 · pp. 5–6
4. Analysis.
The analysis establishes convergence of D3M subdomain networks to variational optima for the Poisson setting and, under stated assumptions, to unique solutions for a class of quasilinear parabolic PDEs. It leaves proof for time-dependent data-free variational formulations to future work.
4. Analysis.
- Under Theorem 1's assumptions, subdomain network iterates converge to the optimal solution of J(ui) once iterations are sufficiently large.The analysis invokes universal approximation in Sobolev space and convergence of Schwarz alternating iterates to the subdomain minimum.Source: §4 · p. 6
- For a fixed boundary function g and q, the optimal solutions of Equations (3.5) and (3.10) satisfy Nu* = v* + g.The paper uses this equivalence to establish that D3M solves steady Poisson’s equation with variational formulations.Source: §4 · p. 7
- Under Assumption 2 and Equation (4.14), D3M networks converge strongly in Lρ(Ωi), for every ρ < 2, to unique solutions of (4.11) as iterations grow.The proof bounds an approximation term by K1ϵ and uses Lipschitz continuity of γ to bound a subsequent term.Source: §4 · pp. 8–9
- The convergence rates to u* are related to the sizes of the overlapping areas between subdomains.Source: §4 · p. 9
- The proof for time-dependent data-free variational formulations is left for future work.Source: §4 · p. 9
5. D3M summary.
D3M divides the domain into components, trains local networks offline, and uses interface information in an online step to estimate target inputs. Its computational cost is dominated by local offline training, and hierarchical neural networks reduce total cost.
5. D3M summary.
- D3M trains local solutions offline and estimates target inputs online, transferring interface information and returning to offline training if solutions do not converge.The offline decomposition avoids data exchange between subdomains; online computation requires no PDE solve.Source: §5 · p. 9
- D3M's PDE computational cost is dominated by local offline training, with total cost expressed using sample size, network blocks, and training times per subdomain.The estimate assumes equal offline sample sizes, block counts, and training epochs across subdomains.Source: §5 · p. 9
- Hierarchical neural networks decrease the total computational cost of D3M.Source: §5 · p. 9
6. Numerical tests.
The tests evaluate D3M on Poisson and time-independent Schrödinger problems, comparing it with DRM, FEM, or analytical solutions. They report higher accuracy than normal DRM in both settings under matched network conditions.
6. Numerical tests.
- The numerical tests use two classical PDE problems and train for 30 epochs with L-BFGS, with reported cost within two minutes on the stated hardware.The hardware is an Intel Core i5-7500, 16GB RAM, and Nvidia GTX 1080Ti.Source: §6 · p. 10
6.1. Poisson’s equation.
- On the square Poisson domain, four overlapping components use random samples that grow by iteration while interface samples stay fixed; residual-block count grows with sample size.The fixed interface sampling provides data exchange; the initial iterations use fewer residual blocks.Source: §6.1 · pp. 10–11, Figure 6.1 · p. 10
- With the same network setting, D3M offers higher accuracy than normal DRM for Poisson’s equation, with FEM used as ground truth.The comparison reports relative errors.Source: Table 6.1 · p. 11, §6.1 · p. 11
6.2. Schrdinger equation.
- For the Schrödinger formulation, Ni approximates the global wave function restricted to subdomain Ωi minus its interface boundary function gi; the cited text considers an infinite potential well.Pi denotes the probability of the particle appearing in Ω\Ωi.Source: §6.2 · p. 12
- For the two-dimensional infinite well, D3M agrees excellently with exact solutions, improves peak predictions over DRM, and has smaller errors for both wave function and probability density under matched conditions.The comparison holds network structure, Lagrangian multiplier, learning rate, sample count, and training epochs constant.Source: §6.2 · p. 13, Table 6.2 · p. 14, Figure 6.5 · p. 14
7. Conclusion.
D3M incorporates domain decomposition into the loss function to support parallel computation, and combines mesh-free sampling with a mixed finite element idea for boundary conditions. The authors identify initialization, interface-function, and Lagrangian-multiplier choices as practical bottlenecks.
7. Conclusion.
- D3M incorporates domain decomposition into the loss function to support parallel computation for large-scale engineering simulations or designs.The conclusion identifies parallel computation as the approach’s most significant contribution.Source: §7 · p. 14
- A mesh-free D3M sampling method is proposed to improve computational efficiency, while a mixed finite element idea aims to satisfy boundary conditions more accurately.Source: §7 · p. 14
- The authors say D3M can solve general parabolic PDEs with high accuracy and suggest its generalization should improve because the domains are smaller.Source: §7 · p. 14
- Three practical bottlenecks remain: poor initialization can add iteration cost, interface-function choice is unresolved, and the Lagrangian multiplier lacks prior guidance.The authors leave these questions for future work.Source: §7 · p. 14