Boundary-Informed Neural Networks

• BINNs are neural architectures that infuse boundary data and constraints into model design, ensuring automatic satisfaction of PDE conditions.
• They employ diverse formulations, including boundary-integral methods, boundary-only inverse models, and structural enforcement approaches, tailored to specific applications.
• Empirical studies show BINNs achieve lower errors and improved efficiency near complex boundaries, though challenges remain in high-frequency and nonlinear regimes.

Searching arXiv for recent and foundational papers on BINNs and related boundary-informed methods. Boundary-Informed Neural Networks (BINNs) denote a family of neural architectures in which boundary data, boundary geometry, boundary operators, or explicit boundary constraints are the primary mechanism by which interior fields, coefficients, or decision surfaces are learned. In the cited literature, the label is not monosemous. It includes “Boundary Integrated Neural Networks” for acoustic radiation and scattering (Qu et al., 2023), boundary-integral type neural networks for computational mechanics (Sun et al., 2023), BINet as a boundary integral network for PDEs (Lin et al., 2021), the “Boundary-Informed Alone Neural Network” (BIAN) for inverse coefficient identification from boundary data only (Chen et al., 14 Jan 2025), and, in a stirred-tank surrogate study, a specific model class with boundary-condition penalties but no PDE residual (Trávníková et al., 15 Jul 2025). A broader boundary-informed reading also encompasses hard boundary-enforced PINN/PINO constructions, latent elliptic boundary-observation operators, and generator-based decision-boundary shaping in supervised learning (Zhou et al., 23 Jul 2025, Göschel et al., 28 Oct 2025, Horsky et al., 2023, Singh et al., 2018).

1. Terminology and conceptual scope

In the cited literature, a BINN is characterized less by a single canonical architecture than by where the physics or geometry is imposed. In one lineage, the governing PDE is rewritten as a boundary integral equation (BIE), all unknowns are transferred to the boundary, and the neural network approximates only boundary traces or boundary densities. In another, the solution ansatz is explicitly factorized so that boundary conditions are satisfied structurally. In a third, boundary observations are encoded and propagated to the interior through a latent elliptic system. A more specialized usage treats “BINN” as a supervised surrogate augmented by boundary-condition residuals but without PDE residuals. A still broader, non-PDE usage treats local decision-boundary geometry as the target of learning via synthetic support points (Chen et al., 14 Jan 2025, Qu et al., 2023, Sun et al., 2023, Trávníková et al., 15 Jul 2025, Singh et al., 2018).

Variant	Defining mechanism	Representative papers
Boundary-integral BINNs	BIE residuals on boundary collocation points	(Lin et al., 2021, Qu et al., 2023, Zhang et al., 2023, Sun et al., 2023, Zhang et al., 28 Feb 2025)
Boundary-only inverse BINNs	Recover interior source/solution/coefficient from boundary measurements only	(Chen et al., 14 Jan 2025)
Boundary-constrained PINN/PINO variants	Distance functions or solution structures enforce BCs structurally	(Zhou et al., 23 Jul 2025, Göschel et al., 28 Oct 2025)
Boundary-observation operator models	Boundary encoder, latent elliptic PDE, decoder	(Horsky et al., 2023)
Data-plus-BC surrogate usage	Data loss plus BC loss, no PDE residual	(Trávníková et al., 15 Jul 2025)
Decision-boundary shaping	Generator creates local support clouds around data	(Singh et al., 2018)

A persistent consequence of this terminological breadth is that “boundary-informed” can refer to fundamentally different mathematical objects: Green’s-function kernels, boundary-only inverse operators, hard boundary ansatzes, or learned neighborhood structure in classification. This suggests that BINNs are best understood as a design principle rather than a single algorithmic template.

2. Boundary-integral formulations as the dominant PDE lineage

The most technically mature BINN lineage rewrites the PDE as a boundary integral problem. In BINet, the solution is represented by a single-layer or double-layer potential,

$$\mathcal{S}[h](x) := -\int_{\partial\Omega} G(x,y)h(y)\,ds_y,\qquad \mathcal{D}[h](x) := -\int_{\partial\Omega} \frac{\partial G(x,y)}{\partial n_y}h(y)\,ds_y,$$

where $G(x,y)$ is the fundamental solution of the operator $\mathcal{L}$. For Dirichlet problems, the boundary density $h$ is learned so that $\mathcal{S}[h]=g$ or $(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$ on $\partial\Omega$. Because $\mathcal{S}[h]$ and $\mathcal{D}[h]$ satisfy $\mathcal{L}u=0$ away from the boundary, training enforces only the boundary condition. This yields automatic PDE satisfaction, dimension reduction from $G(x,y)$0 to $G(x,y)$1, and elimination of differential operators from the optimization loop (Lin et al., 2021).

Acoustic BINNs for the Helmholtz equation use the same principle in direct BIE form. The field satisfies

$G(x,y)$2

where $G(x,y)$3. The network takes only boundary coordinates as input and outputs the real and imaginary parts of the boundary trace. The preferred loss is a single-term BIE residual rather than a composite PDE-plus-BC penalty, which the paper reports as faster and more accurate than the alternative $G(x,y)$4 formulation (Qu et al., 2023).

The same architecture is extended to 2D elastostatics and piezoelectricity by approximating only boundary displacements, tractions, electric potentials, and electric displacements. The governing differential operator is replaced by an integral operator, high-order derivatives of the network are avoided, and all boundary conditions are inherently incorporated within the formulation. This gives a loss consisting solely of BIE residuals, while retaining the classical BEM advantage of boundary-only discretization and validity in bounded and unbounded domains (Zhang et al., 2023).

A computational-mechanics BINN formulation makes the same point in a different language: boundary integral equations are employed to transfer all the unknowns to the boundary, the unknowns are approximated using neural networks, and the loss function is chosen as the residuals of the boundary integral equations. That paper also develops regularization techniques for weakly singular and Cauchy principle integrals, and emphasizes that BINN can easily handle heterogeneous problems with a single neural network without domain decomposition (Sun et al., 2023).

3. Boundary-only inverse problems and the BIAN formulation

BIAN is the most explicit instantiation of a boundary-only inverse BINN. It considers coefficient identification for elliptic PDEs of the form

$G(x,y)$5

with mixed boundary conditions on $G(x,y)$6, and seeks to recover the unknown medium $G(x,y)$7 and the state $G(x,y)$8 only from boundary measurements. The key reformulation writes

$G(x,y)$9

with an equivalent source

$\mathcal{L}$0

Green’s theorem and a weighted residual with Green’s function then produce the boundary–interior relations

$\mathcal{L}$1

for boundary source points and

$\mathcal{L}$2

for interior source points. Boundary values and fluxes thus determine an interior source field through an integral energy/flux balance, without interior measurements (Chen et al., 14 Jan 2025).

The architecture uses three fully connected residual networks: an approximator $\mathcal{L}$3 for the equivalent source $\mathcal{L}$4, a generator $\mathcal{L}$5 for the interior solution $\mathcal{L}$6, and a single-layer discriminator that monitors boundary-error distributions. The approximator is trained by minimizing the residual of the boundary identity; the generator is trained by minimizing the residual of the interior representation; the discriminator computes a Jensen–Shannon divergence between the generator’s boundary error distribution and a uniform distribution, and if $\mathcal{L}$7 the method doubles the number of source points at boundary corners and retrains (Chen et al., 14 Jan 2025).

The theoretical result is boundary-data based. Under space-filling boundary sampling and Hölder continuity assumptions, Theorem 1 yields

$\mathcal{L}$8

with leading behavior

$\mathcal{L}$9

No interior data appear in the probabilistic assumptions. Numerically, for a Laplace problem with spatially varying medium, BIAN reports medium-distribution $h$0 error $h$1 and solution $h$2 error $h$3, compared with $h$4 for PINN, $h$5 for WAN, and $h$6 for DRM. For a Poisson problem with piecewise uniform medium, it reports solution $h$7 error $h$8 and medium $h$9 error $\mathcal{S}[h]=g$0. The paper states that “The derived equivalent source and solution can subsequently be used to infer the indeterminate medium distributions,” but the coefficient-reconstruction step is not as fully detailed as the source/solution learning (Chen et al., 14 Jan 2025).

4. Boundary constraints without boundary integrals

Boundary-informed design is not restricted to BIEs. HB-PINN for Navier–Stokes equations introduces a solution ansatz

$\mathcal{S}[h]=g$1

with a particular solution network $\mathcal{S}[h]=g$2, a distance metric network $\mathcal{S}[h]=g$3, and a primary network $\mathcal{S}[h]=g$4. The distance target is a normalized power-law transform

$\mathcal{S}[h]=g$5

so that $\mathcal{S}[h]=g$6 on the boundary and the interior correction vanishes there. The particular and distance networks are pretrained and frozen; the primary network is then trained only on PDE residuals. This decouples boundary enforcement from interior PDE fitting and yields a structurally boundary-constrained PINN (Zhou et al., 23 Jul 2025).

A closely related but more general operator-level program appears in strong boundary enforcement for PINNs and PINOs. There, weak enforcement penalizes boundary residuals, strong enforcement introduces a solution structure $\mathcal{S}[h]=g$7 that satisfies the boundary condition $\mathcal{S}[h]=g$8 for any differentiable $\mathcal{S}[h]=g$9, and semi-weak enforcement treats only part of the boundary data strongly. Two constructions are proposed: generalized local solution structures (GLSS), which separate distance and normalization functions and handle piecewise $(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$0 but globally $(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$1 boundaries, and orthogonal projections (OP), which use signed distance functions and projections onto boundary hyperplanes. The paper’s central claim is that earlier strong Neumann/Robin methods can become unstable when the boundary is piecewise $(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$2 but only $(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$3 globally, and that GLSS and OP overcome this limitation (Göschel et al., 28 Oct 2025).

SINNs move the boundary-informed principle into a latent operator framework. They learn

$(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$4

where $(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$5 is a boundary encoder, $(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$6 is the solution operator of a latent elliptic system

$(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$7

and $(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$8 is a decoder. The model is trained from pairs $(\frac12\mathcal{I}\pm \mathcal{D})[h]=g$9 without knowledge of the underlying PDE. The latent elliptic system supplies a well-posed boundary-to-interior propagation mechanism even for boundary observation problems that are generic, and even ill-posed (Horsky et al., 2023).

The nomenclature is broader still. In BON, a collaborative generator produces perturbations around each sample so as to create a data support around each original data point which prevents decision boundaries from passing too close to the original data points. BON++ adds Memory Aware Synapses to prevent catastrophic forgetting of these support clouds. Here, “boundary-informed” refers to shaping the classifier’s decision boundary through learned local neighborhoods rather than to PDE boundary conditions (Singh et al., 2018). In the stirred-tank surrogate paper, by contrast, “BINN” is defined operationally as Model III with $\partial\Omega$0, $\partial\Omega$1, $\partial\Omega$2, and $\partial\Omega$3 (Trávníková et al., 15 Jul 2025).

5. Empirical behavior across representative variants

The numerical literature reports strong but highly task-dependent performance. In BIE-based PDE solvers, the typical advantages are boundary-only collocation, smaller networks, and lower errors than PINN baselines. In inverse or hard-constraint variants, the gains are concentrated near complex boundaries and in low-data regimes. Reported headline results include the following.

Representative method	Problem	Reported result
BIAN (Chen et al., 14 Jan 2025)	Laplace with spatially varying medium	Medium $\partial\Omega$4 error $\partial\Omega$5; solution $\partial\Omega$6 error $\partial\Omega$7
Acoustic BINN (Qu et al., 2023)	Interior rectangle, Case 1	Re$\partial\Omega$8 error $\partial\Omega$9; Im$\mathcal{S}[h]$0 error $\mathcal{S}[h]$1
Piezoelectric BINN (Zhang et al., 2023)	Piezoelectric strip	Largest relative error $\mathcal{S}[h]$2; CPU time $\mathcal{S}[h]$3 s
HB-PINN (Zhou et al., 23 Jul 2025)	Cylinder wake flow	$\mathcal{S}[h]$4; $\mathcal{S}[h]$5
Stirred-tank BINN (Trávníková et al., 15 Jul 2025)	$\mathcal{S}[h]$6 labels per $\mathcal{S}[h]$7, $\mathcal{S}[h]$8	$\mathcal{S}[h]$9; $\mathcal{D}[h]$0

In the acoustic Helmholtz study, BINNs achieve relative errors on the order of $\mathcal{D}[h]$1–$\mathcal{D}[h]$2 in several bounded and unbounded examples with 2 hidden layers and about 10–20 neurons, and the pure BIE residual converges much faster than the composite $\mathcal{D}[h]$3 formulation (Qu et al., 2023). In the 2D elastostatic and piezoelectric study, the piezoelectric strip benchmark gives largest relative error $\mathcal{D}[h]$4 and CPU time $\mathcal{D}[h]$5 s for BINN, versus relative errors up to $\mathcal{D}[h]$6 and CPU time $\mathcal{D}[h]$7 s for the compared PINN, on the reported pointwise test (Zhang et al., 2023).

Fracture-mechanics BINNs add Special crack-tip Neural Networks (SPNNs), which encode the asymptotic crack-tip structure directly through

$\mathcal{D}[h]$8

The reported sensitivity study shows that a 2-term SPNN loses accuracy as SPNN size grows, whereas SPNNs with more than 3 terms retain high SIF accuracy even when the SPNN covers $\mathcal{D}[h]$9 of crack length. The same paper reports excellent agreement for mode-I/mode-II fracture parameters and for piezoelectric EDIF/SIF benchmarks in PZT-5H strips (Zhang et al., 28 Feb 2025).

HB-PINN improves boundary-dominated Navier–Stokes benchmarks relative to several PINN baselines. On the cylinder case, the paper reports $\mathcal{L}u=0$0 and $\mathcal{L}u=0$1, versus $\mathcal{L}u=0$2 and $\mathcal{L}u=0$3 for PirateNet and much larger errors for sPINN and hPINN. On the segmented-inlet obstructed cavity, it reports $\mathcal{L}u=0$4 and $\mathcal{L}u=0$5, again outperforming the baselines listed there (Zhou et al., 23 Jul 2025).

The stirred-tank paper places BINNs between PINNs and pure supervised NNs. With $\mathcal{L}u=0$6 labeled points per $\mathcal{L}u=0$7, Model III (“BINN”) reports $\mathcal{L}u=0$8 and $\mathcal{L}u=0$9, better than the pure data NN’s $G(x,y)$00 and $G(x,y)$01, but worse than the PINN-with-data model. With $G(x,y)$02 labels per $G(x,y)$03, BINN and PINN become very similar, while BINN training time is about $G(x,y)$04 minutes versus about $G(x,y)$05 minutes for the PINN-with-data model (Trávníková et al., 15 Jul 2025).

Outside PDEs, BON improves convergence on CIFAR-10 using DenseNet but does not significantly improve final test accuracy, because the generator tends to collapse to the identity mapping. BON++ uses MAS to preserve the synthetic support clouds and yields qualitatively better local boundary shaping on Iris (Singh et al., 2018).

6. Limitations, ambiguities, and active directions

The literature also shows that BINNs are constrained by the mathematical mechanism chosen to inform the network. Boundary-integral BINNs depend on a known fundamental solution, so extension to nonlinear PDEs is not straightforward, and large-scale 3D settings raise the usual BEM difficulties of oscillatory kernels, singular quadrature, and dense operator application. Acoustic BINNs explicitly note that behavior in very high-frequency regimes is not yet explored, while elastostatic and piezoelectric BINNs emphasize that 3D extensions are conceptually straightforward but computationally more demanding (Qu et al., 2023, Zhang et al., 2023).

Hard boundary-enforcement schemes are not universally stable in their naïve forms. Strong enforcement of Neumann or Robin data via earlier distance-function constructions can fail on domains with piecewise $G(x,y)$06 boundaries that are only $G(x,y)$07 globally, which is the rationale for GLSS and OP. This is a genuine methodological controversy: strong enforcement can improve accuracy and training times, but only if the boundary representation is compatible with the regularity of the geometry (Göschel et al., 28 Oct 2025).

Inverse boundary-only BINNs face identifiability and post-processing questions. BIAN provides a fully specified source-learning and solution-learning pipeline and a convergence theorem for the boundary residual, but it does not give full identifiability theorems for $G(x,y)$08, and the coefficient-reconstruction step is less detailed than the equivalent-source and solution stages (Chen et al., 14 Jan 2025).

The broader terminology remains unsettled. The acronym has been used for “Boundary Integrated Neural Networks,” “boundary-integral type neural networks,” “Boundary-Informed Alone Neural Network,” a surrogate with boundary residuals but no PDE residual, and even decision-boundary support generation in supervised learning. This suggests that the stable content of the field lies not in the acronym itself but in a recurring architectural principle: boundary information is lifted from an auxiliary penalty to a primary structural component of the model (Chen et al., 14 Jan 2025, Qu et al., 2023, Trávníková et al., 15 Jul 2025, Singh et al., 2018).

The active directions named in the cited papers are correspondingly diverse: structural–acoustic sensitivity analysis, 3D elastostatics and 3D piezoelectricity, fast multipole and hierarchical-matrix acceleration, adaptive boundary sampling, hybrid boundary/interior schemes, iterative linearization for weakly nonlinear problems, and extensions to more complex coupled systems (Qu et al., 2023, Zhang et al., 2023, Zhou et al., 23 Jul 2025, Göschel et al., 28 Oct 2025). Across these variants, the unifying thesis is consistent: when the relevant operator, constraint, or data-acquisition process is fundamentally boundary-driven, a neural architecture that encodes that fact explicitly can reduce sampling burden, sharpen physical fidelity near boundaries, and alter the conditioning of the learning problem itself.