Universal Subdomain Solver (USDS)

• USDS is a method that reuses one CNN to solve local PDE subdomains by encoding geometry via signed distance functions and boundary conditions.
• It integrates a physics-aware hard constraint layer to enforce flow-rate conservation within overlapping Schwarz iterations, ensuring global solution consistency.
• USDS demonstrates adaptability in various settings, including transfer learning and nonstationary PDEs, highlighting its potential to unify local surrogate modeling.

Searching arXiv for the cited USDS paper and a few related subdomain/domain-decomposition works to ground the article. Universal Subdomain Solver (USDS) denotes a single learned or algorithmic solver that is reused across multiple subdomains rather than being specialized to one fixed subproblem. In the most specific and technically developed sense documented here, the term refers to the convolutional-neural-network-based subdomain solver introduced for overlapping Schwarz iterations in non-Newtonian blood-flow simulation, where one CNN is trained once on local artery patches and then deployed unchanged on every subdomain across many geometries and inflow conditions (Klaes et al., 19 Sep 2025). More broadly, the term also appears as a useful organizing concept for subdomain-aware methods in transfer learning, distribution shift, and nonstationary PDE decomposition, where “subdomain” may refer to spatial patches, low-dimensional manifolds, or support-defined regions rather than geometric subregions alone (Wei et al., 2020, Fang et al., 2023, Vabishchevich, 2022). Across these settings, the unifying idea is reuse of a common local mechanism together with an explicit scheme for coupling subdomains into a global solution.

1. Definition and scope

In the blood-flow formulation of CNN-Schwarz-Flow, a Universal Subdomain Solver is a single convolutional neural network that acts as an inexact PDE solver on every subdomain in an overlapping Schwarz domain decomposition scheme (Klaes et al., 19 Sep 2025). The global two-dimensional artery domain $\Omega$ is decomposed into rectangular patches on a pixel grid, and one CNN is trained on many such local patches extracted from finite-element simulations on a fixed arterial setup. The network maps local geometry information, encoded by a signed distance function, together with local boundary velocity data, and in the constrained variant the inlet flow rate, to the velocity field on that patch. After training, the same model is used for all subdomain solves in the alternating Schwarz iteration, across different arteries and inflow conditions, without retraining (Klaes et al., 19 Sep 2025).

The paper contrasts this with both classical domain decomposition and naive surrogate modeling. In classical domain decomposition, each subdomain problem is solved numerically, for example by the finite element method, with exact or controlled accuracy. A naive data-driven alternative would either train separate networks for different geometries or subdomain types, or train one global network that maps global boundary data directly to a global field on a fixed domain. Those alternatives lack transferability when geometry length, stenosis location, stenosis shape, or boundary conditions change. The USDS instead learns a local solution operator conditioned on patch geometry and interface data, making reuse across arbitrarily many subdomains and longer arteries feasible within the same family of problems (Klaes et al., 19 Sep 2025).

Outside geometric PDE decomposition, the phrase is also useful in a more abstract sense. In unsupervised domain adaptation, subdomains can be defined as low-dimensional manifolds, and a universal subdomain-oriented mechanism would jointly discover those manifolds and minimize a local discrepancy on each of them rather than only a global domain discrepancy (Wei et al., 2020). In distribution-shift learning, a universal solver is realized by splitting test support into in-training and out-of-training regions and assigning different objectives to the two regions while preserving global risk consistency (Fang et al., 2023). In nonstationary PDEs, solution decomposition methods define subdomain components through restriction and extension operators and then apply generic subdomain time-marching schemes derived from the block operator structure (Vabishchevich, 2022). These usages are not identical, but they share a structural motif: a fixed local mechanism is embedded inside a coupling framework that produces a global solution.

2. Role in CNN-Schwarz-Flow for blood-flow simulation

The most concrete USDS instantiation appears in “A Flow-rate-conserving CNN-based Domain Decomposition Method for Blood Flow Simulations” (Klaes et al., 19 Sep 2025). Blood is modeled as a stationary, incompressible, shear-thinning, generalized Newtonian fluid in a rigid two-dimensional artery. The governing equations are mass and momentum conservation,

$$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$

$$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$

with Dirichlet conditions at the inlet and walls, and a stress-free Neumann condition at the outlet. The constitutive law is generalized Newtonian with shear-rate-dependent viscosity $\eta(\dot{\gamma})$, implemented through a Carreau-type model fitted to rheometer data (Klaes et al., 19 Sep 2025).

The artery is represented on a pixel grid of height $H=128$, with global width depending on artery length, and decomposed into overlapping rectangular subdomains $I_n$ of fixed size $256 \times 128$ pixels. Neighboring subdomains overlap in the axial direction. The global problem is then solved by an alternating overlapping Schwarz method. In each red phase, the USDS is called on even-index subdomains using the signed distance function and boundary velocity data taken from the current iterate; in each black phase, the same is done for odd-index subdomains. Iteration stops when the maximum pointwise update satisfies

$$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$

Because the CNN is an inexact local solver, classical Schwarz convergence guarantees do not apply directly, and convergence is studied empirically rather than established by standard theory (Klaes et al., 19 Sep 2025).

This embedding of a learned local operator inside a classical domain-decomposition loop is central. The USDS does not replace the global iterative structure; it replaces the local PDE solve. A plausible implication is that the framework separates two sources of generalization: local generalization within the CNN and global consistency induced by Schwarz iteration. The paper itself emphasizes that these interact strongly, especially when the local solver is not explicitly constrained by physics (Klaes et al., 19 Sep 2025).

3. Mathematical formulation of the learned subdomain operator

The data-driven USDS is defined as a CNN mapping

$$\mathcal{N}_\theta : \mathbb{R}^{3 \times W \times H} \to \mathbb{R}^{2 \times W \times H},$$

$$\mathcal{N}_\theta \big(\text{SDF}, v_x(\xi), v_y(\xi)\big) = (\hat{v}_x, \hat{v}_y),$$

where the first input channel is the signed distance function and the second and third channels encode the boundary velocities in boundary strips of width $$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$0 at the left and right patch edges (Klaes et al., 19 Sep 2025). The output is the velocity field on the full subdomain.

Two post-processing steps are part of the operator. First, a no-slip mask is defined from the signed distance function,

$$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$1

and the predicted velocity is overwritten by $$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$2, enforcing zero velocity outside the fluid region and at walls. Second, the boundary data in the strip $$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$3 are overwritten exactly so that the patch respects the supplied interface values (Klaes et al., 19 Sep 2025).

The physics-aware, flow-rate-conserving variant augments this architecture with a hard constraint layer. In a channel aligned with the $$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$4-direction, the flow rate across a cross-section is

$$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$5

with discrete approximation

$$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$6

For incompressible stationary flow without sources or sinks, the desired condition is $$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$7 for all columns. The raw CNN prediction $$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$8 is corrected columnwise by

$$\rho (\mathbf{v} \cdot \nabla)\mathbf{v} + \nabla p - \nabla \cdot \boldsymbol{\tau} = \mathbf{f} = 0 \quad \text{in } \Omega,$$9

which enforces exact cross-sectional flow-rate conservation,

$$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$0

The corresponding operator is

$$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$1

$$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$2

The architecture otherwise remains the same bottleneck CNN with shared encoder and two decoders (Klaes et al., 19 Sep 2025).

A notable design point is that the physics awareness is architectural rather than loss-based. The training objective is plain mean squared error,

$$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$3

and the paper explicitly states that there is no explicit physics loss term; the physics comes from the hard constraint layer and from the mask and boundary overwrite (Klaes et al., 19 Sep 2025). This distinguishes the method from more diffuse uses of “physics-informed” training.

4. Training strategy and the meaning of universality

The training corpus is generated from 770 finite-element simulations on stenotic artery geometries with a fixed global configuration: maximum diameter $$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$4, total length $$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$5, inlet and outlet lengths $$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$6 each, and a $$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$7 stenotic region with up to three segments of random length and severity (Klaes et al., 19 Sep 2025). Geometric variation is introduced through the number of stenotic segments, relative segment lengths, and upper and lower stenosis factors. Inflow variation is introduced through $$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$8. The finite-element reference uses P2–P1 elements, Picard plus Newton linearization, and an overlapping Schwarz preconditioner (Klaes et al., 19 Sep 2025).

Each global solution is sampled to a $$\nabla \cdot \mathbf{v} = 0 \quad \text{in } \Omega,$$9 pixel grid and decomposed into local patches of size $\eta(\dot{\gamma})$0. The dataset contains 9 non-overlapping subdomains per simulation plus 20 additional overlapping subdomains in the stenotic region, for a total of $\eta(\dot{\gamma})$1 subdomain images (Klaes et al., 19 Sep 2025). The signed distance function is normalized to $\eta(\dot{\gamma})$2 using max normalization, while velocities are left unnormalized because of scale differences and numerical artifacts. Training uses an 80/10/10 train/validation/test split, Adam with learning rate $\eta(\dot{\gamma})$3, batch sizes 64 and 32 for training and validation/test respectively, early stopping with patience 50 and minimum validation-loss improvement $\eta(\dot{\gamma})$4, and distributed Horovod training on GPUs, taking about two hours for 400 epochs (Klaes et al., 19 Sep 2025).

Within this framework, “universal” has a specific operational meaning. Every patch has the same dimensions $\eta(\dot{\gamma})$5 and covers one full artery cross-section, so the network architecture can be reused on any subdomain. The CNN sees geometry only through the signed distance function and interface conditions only through boundary strips, which abstracts away global position. The model is trained once on local patches drawn from one consistent arterial setup and is then reused for any decomposition of any artery in that family, including arbitrarily long arteries composed of more patches (Klaes et al., 19 Sep 2025).

The following summary captures the main ingredients of universality in this sense.

Ingredient	Description	Function
Fixed patch shape	$\eta(\dot{\gamma})$6 pixels	Reusable CNN input/output format
Geometry encoding	Signed distance function	Position-independent local geometry representation
Interface encoding	Boundary velocity strips	Local transmission of Schwarz boundary data
Single training stage	One model trained on pooled patches	Reuse across subdomains and future geometries

This universality is therefore conditional rather than absolute. It is universal across subdomains and across many arteries within a structured family, not across arbitrary PDE classes or arbitrary vascular geometries. The paper itself notes that the approach is limited to two-dimensional arteries, a specific Carreau-type viscosity model, rigid walls, and stationary flow (Klaes et al., 19 Sep 2025).

5. Flow-rate conservation, convergence behavior, and validation

The central empirical result is that a physics-aware USDS preserving flow rate is necessary when the amount of training data is limited (Klaes et al., 19 Sep 2025). The motivation is straightforward. In a Schwarz method with inexact local solvers, if subdomain predictions do not conserve mass flux, successive subdomain updates can accumulate either undershooting, in which the flow rate decays along the artery, or overshooting, in which it grows unphysically. The paper reports that such errors degrade or destroy global convergence: the iteration may converge to an invalid field or diverge (Klaes et al., 19 Sep 2025).

Global accuracy is measured by the Global Relative Error,

$\eta(\dot{\gamma})$7

with analogous GRE$\eta(\dot{\gamma})$8 and GRE$\eta(\dot{\gamma})$9 for the components, evaluated only inside the fluid region (Klaes et al., 19 Sep 2025). Results are classified into bins $H=128$0, $H=128$1, $H=128$2, $H=128$3, $H=128$4, and diverged.

For the flow-rate-conserving USDS, the majority of predictions have GRE below $H=128$5 for $H=128$6, with $H=128$7 giving the best overall performance: about $H=128$8 of cases have GRE below $H=128$9 (Klaes et al., 19 Sep 2025). For cases with GRE $I_n$0, the averages are approximately $I_n$1 for GRE, $I_n$2 for GRE$I_n$3, and $I_n$4 for GRE$I_n$5, reflecting the comparatively larger relative error in the smaller transverse velocity component. Convergence typically requires fewer than 25–30 red–black cycles (Klaes et al., 19 Sep 2025).

The purely data-driven USDS performs markedly worse. Its best case, again at $I_n$6, yields about $I_n$7 of predictions with GRE below $I_n$8, and larger $I_n$9 increases divergence and poor accuracy because more interior error is frozen into the boundary strips and propagated iteratively (Klaes et al., 19 Sep 2025). The reported qualitative failure modes are vanishing flow due to mass loss and blow-up in $256 \times 128$0 contaminating $256 \times 128$1. The constrained USDS is not immune to divergence, because the $256 \times 128$2 component remains unconstrained, but divergence becomes much less frequent and is concentrated in severe stenoses with high inflow and high shear (Klaes et al., 19 Sep 2025).

Scalability tests reinforce the local-to-global claim. When the stenotic region is extended by factors 2, 4, and 8, using duplicated or random stenotic patterns, the physics-aware USDS retains small GRE and converges in roughly 20–30 red–black iterations. The paper reports that convergence depends more strongly on local flow complexity than on the number of subdomains, and suggests that interior subdomains still receive global information through $256 \times 128$3 (Klaes et al., 19 Sep 2025). This suggests that the hard flux constraint acts as a low-dimensional global coupling variable inside an otherwise local solver.

6. Relations to other meanings of subdomain solving

The term USDS is not standardized across fields, and related work uses “subdomain” in materially different senses. In transfer learning, “Subdomain Adaptation with Manifolds Discrepancy Alignment” defines subdomains as low-dimensional manifolds shared by source and target domains and proposes Manifold Maximum Mean Discrepancy (M3D) as a local discrepancy measure (Wei et al., 2020). Rather than solving PDEs, TMDA minimizes

$256 \times 128$4

where $256 \times 128$5 is a manifold-discovery term, $256 \times 128$6 is the manifold discrepancy term, and $256 \times 128$7 is regularization (Wei et al., 2020). Here the subdomain concept is latent and statistical rather than geometric. The connection to USDS is conceptual: a universal subdomain-oriented method should discover local structure and solve local mismatch rather than treating the problem as globally homogeneous.

In distribution shift, “Generalizing Importance Weighting to A Universal Solver for Distribution Shift Problems” provides another abstract template (Fang et al., 2023). The method splits the test support into an in-training region $256 \times 128$8 and an out-of-training region $256 \times 128$9, then defines the generalized importance-weighting objective

$$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$0

The paper proves that this objective is risk-consistent in all four support-overlap cases and reduces to classical importance weighting when $$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$1 (Fang et al., 2023). Here the “universal solver” is universal across support relations because it assigns different local objectives to different support-defined subdomains while preserving the correct global risk.

In nonstationary PDEs, “Subdomain solution decomposition method for nonstationary problems” studies a different form of universality (Vabishchevich, 2022). The solution is decomposed as

$$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$2

leading to a block operator

$$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$3

and subdomain time-marching schemes based on diagonal or triangular splitting (Vabishchevich, 2022). This is not a learned solver, but it provides a rigorous template in which a single subdomain-level construction can be applied across overlapping and non-overlapping decompositions and across a class of linear parabolic problems with self-adjoint positive operators. A plausible implication is that learned USDS methods can be interpreted as replacing the local action of $$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$4 while retaining classical operator-level coupling.

These three lines of work show that “subdomain” can mean spatial patch, manifold, or support region, and that “universal” can mean reusable across geometry instances, transferable across latent local structures, or risk-consistent across support-overlap regimes. They should not be conflated, but they illuminate a common design logic: decompose globally, solve locally with a shared mechanism, then recouple.

7. Limitations, misconceptions, and prospective extensions

A common misconception is that universality implies unrestricted generalization. In the blood-flow setting, the USDS is not universal across arbitrary geometries, dimensions, or physics. It is trained on one fixed arterial configuration with prescribed diameter and inlet/outlet structure, then reused on local patches from arteries in that family (Klaes et al., 19 Sep 2025). The paper explicitly identifies several limitations: only two-dimensional arteries are considered; only flow-rate conservation in the $$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$5-direction is enforced; full continuity $$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$6 is not enforced; the viscosity model is a specific generalized Newtonian Carreau-type law; and the flow is stationary with rigid walls (Klaes et al., 19 Sep 2025).

Another misconception is that better local patch accuracy automatically implies better global Schwarz performance. The results show this is false. Increasing the boundary strip width $$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$7 improves local mean squared error for both architectures, but the globally best performance occurs at $$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$8, whereas $$\text{AbsErr}_{\max} = \max_{i,j} \left| v^{k+1}(i,j) - v^k(i,j) \right| < \varepsilon.$$9 gives the smallest subdomain MSE yet worsens global predictions and causes some divergence (Klaes et al., 19 Sep 2025). The paper attributes this to iterative propagation of errors frozen into the enlarged boundary strips. This demonstrates that a USDS must be evaluated as part of the coupled algorithm, not only as an isolated local predictor.

The work also clarifies a distinction between physics-aware and fully physics-informed modeling. The flow-rate-conserving variant enforces a hard mass-flux condition but does not enforce the full incompressibility constraint or solve for pressure (Klaes et al., 19 Sep 2025). The remaining instability through $$\mathcal{N}_\theta : \mathbb{R}^{3 \times W \times H} \to \mathbb{R}^{2 \times W \times H},$$0 in severe stenoses makes this limitation empirically visible. This suggests, as the authors note, extensions involving hard or soft enforcement of full continuity, integration of pressure, and more expressive architectures such as neural operators, U-Net-like models, or graph-based methods (Klaes et al., 19 Sep 2025).

More broadly, the blood-flow USDS suggests several transferable design principles that resonate with the other cited works. Patch-based learning with geometry encoded independently of global position parallels manifold-based local representation in TMDA (Wei et al., 2020). Hard architectural constraints that preserve key invariants are analogous to GIW’s exact decomposition of risk into valid local objectives across support-defined regions (Fang et al., 2023). Embedding a learned local solver inside a principled coupling scheme mirrors the operator-theoretic role of restriction, extension, and block coupling in solution decomposition methods for nonstationary PDEs (Vabishchevich, 2022).

Future work identified in the blood-flow study includes extension to three-dimensional arteries, curved vessels, and branching networks; enforcement of full continuity; integration of pressure and additional quantities; exploration of improved architectures and losses; multi-fidelity or reduced-order augmentation; and theoretical analysis of Schwarz convergence with inexact neural subdomain solvers (Klaes et al., 19 Sep 2025). This suggests that the long-term significance of the USDS concept lies less in one particular CNN architecture than in a broader methodological program: constructing scalable local surrogate operators that can be inserted into classical decomposition frameworks without sacrificing the physical or statistical invariants required for reliable global behavior.