Physics-Inspired Domain Decomposition

• Physics-Inspired Domain Decomposition is a method for partitioning complex PDE domains into subdomains where tailored neural surrogates enforce underlying physics.
• It leverages overlapping and non-overlapping partitioning, adaptive multilevel architectures, and physics-guided interface coupling to mitigate issues like spectral bias and communication bottlenecks.
• The approach delivers significant error reductions (up to 1–3 orders of magnitude) and near-linear parallel scaling, making it effective for large-scale, multi-physics problems.

Physics-Inspired Domain Decomposition

Physics-inspired domain decomposition refers to a family of methodologies for decomposing a complex physical domain—typically arising in partial differential equation (PDE) modeling—into smaller subdomains, then constructing, coupling, and training neural- or operator-based surrogates that leverage the underlying physics to achieve scalable and accurate solution of the governing equations. This concept is pivotal in overcoming critical limitations of global physics-informed neural networks (PINNs), particularly for large-scale, multi-scale, or multi-physics problems, and underlies diverse approaches ranging from symmetry-based decomposition to hybrid neural–operator techniques (Liu et al., 2024).

1. Theoretical Foundations and Motivation

Classical numerical methods, such as finite element or finite difference methods, exploit domain decomposition to enable parallelism, handle complex geometries, and address multi-scale phenomena. Their physics-driven subdomain coupling principles (e.g., transmission operators in Schwarz-type methods) are now being adapted and extended to neural approximators, with critical modifications to account for the optimization properties and inductive biases inherent in machine learning-based surrogates.

Neural network-based domain decomposition methods address: (i) spectral bias in learning high-frequency or multi-scale solutions; (ii) prohibitive parameter counts for large domains; (iii) communication bottlenecks in distributed computing; (iv) robustness in noisy or heterogeneous data settings; and (v) the need for "train once, solve many" generalization, particularly in neural operators (Wu et al., 23 Jul 2025).

2. Core Methodological Components

2.1. Subdomain Partitioning Strategies

Domain decomposition schemes can be grouped as overlapping (e.g., Schwarz or partition of unity) or non-overlapping (e.g., conformal mesh, non-overlapping Schwarz-type). Overlapping strategies employ smooth window functions—such as cosine tapers (Howard et al., 2024) or sigmoidal partitions (Moseley et al., 2021)—to ensure global $C^0$ or higher-order continuity and natural blending of subdomain solutions. Non-overlapping schemes rely on explicit interface enforcement, often through equality-constrained losses with learnable Robin or hybrid transmission parameters (Basir et al., 2023, Hu et al., 2024).

A general global ansatz for solution $u$ in an overlapping setting is

$u(x) = \sum_{i=1}^N \omega_i(x)u_i(x;\theta_i)$

with $\sum_i \omega_i(x) \equiv 1$ and each $u_i$ a subdomain-local neural network (Moseley et al., 2021).

2.2. Physics-Guided Interface and Coupling

Physics-inspiration enters primarily through interface treatment. Conservative quantities (e.g., fluxes in elliptic/hyperbolic problems) are enforced via explicit flux continuity constraints (Figueres et al., 26 Apr 2025, Shukla et al., 2021), generalized Robin-type conditions (Basir et al., 2023, Hu et al., 2024), or through weak-form projections in variational PINNs (VPINNs) (Kharazmi et al., 2020). In overlapping settings, smooth blending via partition of unity and local input normalization implicitly enforce continuity; in non-overlapping settings, interface condition optimization is essential for global accuracy and information propagation.

Symmetry-based domain decomposition extends this principle: interfaces are constructed as invariant manifolds for a PDE's Lie symmetry group (sdPINN), ensuring that interface data are "exact" and physically consistent (Liu et al., 2024).

2.3. Local and Global Loss Design

Each subdomain network is trained using physics-informed losses composed of residuals for the governing PDE, boundary/initial conditions, and interface/overlap penalties:

• In overlapping settings, global residuals are evaluated with the partition-of-unity-composed surrogates, and additional interface continuity losses may be used optionally.
• In non-overlapping or hybrid cases, hard or augmented Lagrangian constraints are used at interfaces, often with learnable transmission parameters to optimize convergence (Basir et al., 2023, Hu et al., 2024).

Multilevel and multifidelity extensions include the addition of coarse-level "global" neural networks—paralleling multigrid or Schwarz two-level methods—to propagate low-frequency information throughout the domain and address the convergence slowdown ("Schwarz bottleneck") observed in fine-level-only decompositions (Dolean et al., 2024, Dolean et al., 2023).

2.4. Parallelism, Optimization, and Scalability

These methodologies naturally admit distributed implementation: all subdomain networks (and, in multilevel schemes, the coarse network) can be trained or evaluated in parallel, with communication localized to overlap or interface regions. Modern approaches exploit this to attain near-linear weak scaling on multi-GPU or multi-node clusters in both training and inference (Feeney et al., 2023, Wu et al., 23 Jul 2025).

Iterative solvers such as L-BFGS, Adam, and (block-sparse) Gauss–Newton are used per subdomain, with interface data exchanged at synchronization points. Here, block-sparsity of the overall system's Jacobian induced by domain decomposition is explicitly exploited to accelerate convergence and reduce computational cost (Heinlein et al., 30 Oct 2025).

Bayesian domain decomposition methods further allow each subdomain to carry an independent uncertainty quantification, with global posteriors assembled from local predictions (Figueres et al., 26 Apr 2025).

3. Symmetry-Group and Physics-Driven Decompositions

Recent methodology advances include the use of Lie-symmetry group actions for generating physically meaningful interfaces in PINN-based solvers:

• A Lie-symmetry group $G_\epsilon$ induces diffeomorphisms, whose orbits are parameterized by invariants $I_1(x,t,u), I_2(x,t,u)$. By selecting seeds on initial or boundary data and propagating group orbits, one generates interfaces that carry invariant solution information, divides the domain into subdomains, and guarantees exact, physics-respecting interface data (Liu et al., 2024).
• On each subdomain, PINNs or symmetry-enhanced PINNs are trained. Losses incorporate both standard physics-informed residuals and "invariant surface conditions" (e.g., $g(x, t; u) := \xi u_x + \tau u_t - \eta = 0$) tied to the infinitesimal symmetry.
• Numerical results for the Korteweg-de Vries and nonlinear viscous fluid equations show that this approach outperforms both vanilla PINNs and XPINN in terms of mean $L_2$-relative errors—by 1–2 orders of magnitude—with extreme stability and accuracy at interfaces (Liu et al., 2024).

4. Advanced and Adaptive Decomposition Architectures

4.1. Adaptive and Residual-Driven Decomposition

Residual-driven adaptive-basis PINNs (AB-PINNs) introduce a dynamic domain decomposition, spawning new subdomains in regions with persistently high PDE residuals:

• Local basis functions (e.g., windowed RBFs) adapt their centers and shapes via gradient descent, clustering subdomain "experts" in regions of maximum physics-induced complexity.
• As new subdomains are added, spikes in the residual are rapidly diminished, yielding 1–3 orders of magnitude improvements in $L^2$ errors over static decompositions, particularly in multiscale or localized features (Botvinick-Greenhouse et al., 10 Oct 2025).

4.2. Multilevel and Multiscale Architectures

Multilevel FBPINNs combine hierarchical overlapping decompositions, where coarse levels capture global, low-frequency features and fine levels resolve high-frequency, localized physics (Dolean et al., 2023, Dolean et al., 2024):

• These architectures show strong scaling—error decreases as model capacity increases, and weak scaling—error remains roughly constant when problem complexity and capacity grow together.
• In high-frequency Helmholtz or multi-frequency Poisson problems, multilevel strategies enable accurate solution recovery with computational costs scaling only with the number of overlapping patches, not the global frequency (Dolean et al., 19 Nov 2025).

4.3. Operator-Learning and Pretrained Solvers

Neural operator-based domain decomposition (e.g., L-DDM) introduces pre-trained physics-informed neural operators as black-box subdomain solvers within an additive Schwarz framework:

• A single neural operator, trained on canonical (small) domains, is reused in the online phase across arbitrarily large or complex-patched domains (Wu et al., 23 Jul 2025).
• Global assembly via the classical Schwarz iterations retains convergence guarantees, while attaining resolution-invariance and generalization to novel microstructural or boundary conditions.
• This approach achieves net performance (in relative error and wall time) competitive with or superior to state-of-the-art numerical and operator-based methods.

5. Practical Performance and Benchmarking

Comprehensive empirical studies consistently demonstrate:

• Order-of-magnitude reductions in $u$0 or $u$1 errors for FBPINNs and variants versus single-network PINNs, even at high frequencies or in domains with sharp gradients (Moseley et al., 2021, Saha et al., 2024, Dolean et al., 19 Nov 2025).
• Robustness to data noise, with Bayesian decomposed PINNs yielding credible intervals within 5–15% noise and outperforming single-domain Bayesian PINNs in global uncertainty quantification (Figueres et al., 26 Apr 2025).
• Strong and weak parallel scaling—Mosaic Flow, XPINN/cPINN, and L-DDM achieve near-linear scaling on up to 32 GPUs or clusters of CPUs, with only minimal communication overhead (Feeney et al., 2023, Shukla et al., 2021).
• Multilevel and adaptive methods outperform single-level decompositions both in accuracy and wall time at large numbers of subdomains, with negligible overhead from global (coarse) network updates (Dolean et al., 2024).

Method	$u$2 Error	Parallel Scalability	Notable Advantage
sdPINN (symmetry)	$u$3–$u$4 (vs $u$5–$u$6)	Yes	Interface exactness, parallelism (Liu et al., 2024)
FBPINN	$u$7–$u$8 (vs $u$9–$u(x) = \sum_{i=1}^N \omega_i(x)u_i(x;\theta_i)$0)	Yes	Spectral-bias mitigation (Moseley et al., 2021, Dolean et al., 19 Nov 2025)
Mosaic Flow	$u(x) = \sum_{i=1}^N \omega_i(x)u_i(x;\theta_i)$1	$u(x) = \sum_{i=1}^N \omega_i(x)u_i(x;\theta_i)$285% on 32 GPUs	Reusable subnet, no retrain (Feeney et al., 2023)

6. Extensions, Limitations, and Ongoing Research

Physics-inspired domain decomposition generalizes readily to operator learning, uncertainty quantification, multifidelity modeling, and complex geometries:

• Multifidelity stacking and deep operator networks benefit from time-domain decomposition, with multifidelity FBPINNs achieving sub-1% relative errors on challenging benchmarks (Heinlein et al., 2024).
• Learning or optimizing interface transmission parameters (Robin, tangential, etc.) augments classical Schwarz methods and adapts to heterogeneous local physics (Basir et al., 2023, Hu et al., 2024).
• Adaptive spawning and pruning of subdomains, guided by PDE-residual or physics-based indicators, offer enhanced solution adaptivity for time-varying or regionally dynamic problems (Botvinick-Greenhouse et al., 10 Oct 2025).
• Challenges remain in optimal overlap sizing, interface penalty weight selection, and efficient inversion or approximation of per-level block-sparse Hessians, particularly for high-dimensional or real-time settings (Dolean et al., 2023, Heinlein et al., 30 Oct 2025).

7. Historical Context and Impact

The development of physics-inspired domain decomposition in machine learning emerges from integration of classical domain decomposition (Schwarz-type, optimized, and multilevel) with modern neural approximators and operator frameworks. Contemporary advances—ranging from symmetry-driven interfaces (Liu et al., 2024) to multilevel, multifidelity, and Bayesian uncertainty approaches—demonstrate a trend toward highly parallel, adaptive, physics-respecting, and scalable machine learning solvers, fundamentally extending the frontier of what neural-network-based surrogates can achieve in scientific computing.

---

References:

• "Symmetry group based domain decomposition to enhance physics-informed neural networks for solving partial differential equations" (Liu et al., 2024)
• "Breaking Boundaries: Distributed Domain Decomposition with Scalable Physics-Informed Neural PDE Solvers" (Feeney et al., 2023)
• "hp-VPINNs: Variational Physics-Informed Neural Networks With Domain Decomposition" (Kharazmi et al., 2020)
• "$PINN - a Domain Decomposition Method for Bayesian Physics-Informed Neural Networks" (Figueres et al., 26 Apr 2025)
• "Towards Model Discovery Using Domain Decomposition and PINNs" (Saha et al., 2024)
• "Initialization-enhanced Physics-Informed Neural Network with Domain Decomposition" (Si et al., 2024)
• "Multilevel domain decomposition-based architectures for physics-informed neural networks" (Dolean et al., 2023)
• "Finite Basis Physics-Informed Neural Networks (FBPINNs): a scalable domain decomposition approach for solving differential equations" (Moseley et al., 2021)
• "Domain decomposition architectures and Gauss-Newton training for physics-informed neural networks" (Heinlein et al., 30 Oct 2025)
• "Non-overlapping, Schwarz-type Domain Decomposition Method for Physics and Equality Constrained Artificial Neural Networks" (Hu et al., 2024)
• "AB-PINNs: Adaptive-Basis Physics-Informed Neural Networks for Residual-Driven Domain Decomposition" (Botvinick-Greenhouse et al., 10 Oct 2025)
• "Parallel Physics-Informed Neural Networks via Domain Decomposition" (Shukla et al., 2021)
• "A Generalized Schwarz-type Non-overlapping Domain Decomposition Method using Physics-constrained Neural Networks" (Basir et al., 2023)
• "The modified Physics-Informed Hybrid Parallel Kolmogorov--Arnold and Multilayer Perceptron Architecture with domain decomposition" (Huang et al., 14 Nov 2025)
• "Finite basis Kolmogorov-Arnold networks: domain decomposition for data-driven and physics-informed problems" (Howard et al., 2024)
• "A Learning-based Domain Decomposition Method" (Wu et al., 23 Jul 2025)
• "Neural network-driven domain decomposition for efficient solutions to the Helmholtz equation" (Dolean et al., 19 Nov 2025)
• "Two-level deep domain decomposition method" (Dolean et al., 2024)
• "Multifidelity domain decomposition-based physics-informed neural networks and operators for time-dependent problems" (Heinlein et al., 2024)