Boundary Enforcing Operator Network (BOON)

• BOON is a class of neural operator architectures that enforces exact boundary conditions by integrating specialized solution-structure layers and kernel correction techniques.
• It leverages methods like GLSS and Orthogonal-Projection to transform traditional soft BC constraints into strong, exact enforcement suitable for Dirichlet, Neumann, Robin, and periodic conditions.
• Empirical findings demonstrate that BOON improves boundary fidelity, accelerates convergence, and reduces computational overhead compared to penalty-based approaches.

Boundary Enforcing Operator Network (BOON) is a class of neural operator architectures and kernel correction techniques designed to enforce boundary conditions (BCs) of partial differential equations (PDEs) exactly within machine learning-based solution frameworks. BOONs ensure physical consistency by integrating structure-preserving layers or kernel modifications, enabling neural operators to satisfy Dirichlet, Neumann, Robin, or periodic BCs by construction rather than by incorporating penalty terms into the loss function.

1. Theoretical Foundations and Motivation

Neural operators, including variants such as Fourier Neural Operator (FNO), DeepONet, and Multiwavelet Operator (MGNO), are widely used for learning input-output maps of PDEs in arbitrary geometries. However, standard implementations typically incorporate BCs as soft constraints using boundary loss penalties, leading to solutions that may not strictly adhere to physical requirements. The absence of exact BC satisfaction can result in non-physical predictions, compromised uniqueness, and suboptimal global accuracy, particularly for stiff or sensitive problems (Göschel et al., 28 Oct 2025, Saad et al., 2022, Wu et al., 16 Jan 2026).

BOON methodologies address this limitation by recasting the neural operator or its kernel (discrete or continuous) to ensure that the network output matches prescribed BCs by construction. The BOON paradigm includes both operator kernel correction approaches and solution-structure layers embedded in neural architectures.

2. Mathematical Formulations and Enforcement Mechanisms

Three principal BOON variants have been established:

2.1 Solution-Structure Layers for Physics-Informed Neural Operators

Two strong enforcement techniques have been introduced in (Göschel et al., 28 Oct 2025):

• Generalized Local Solution-Structure (GLSS) Method: This formulation applies to domains with boundaries partitioned into $M$ $C^1$-smooth segments, each carrying Dirichlet or Robin/Neumann BCs. For each segment $\Gamma_i$, auxiliary distance functions $\phi_i(x)$ vanishing on $\Gamma_i$ and normalized functions $\bar{\phi}_i(x)$ are used to construct a transfinite interpolant:

$$u(x) = \sum_{i=1}^M w_i(x)\, u_i(x) + \Psi_{\text{rem}}(x)\prod_{i=1}^M \phi_i(x)^{\mu_i}$$

where $\mu_i=1$ for Dirichlet and $\mu_i=2$ for Robin/Neumann. The local structures $u_i$ are carefully defined to encode exact boundary behavior, including Taylor-R-function constructions for Robin/Neumann and explicit correction terms to maintain regularity at corners. Piecewise-$C^1$ intersection corrections guarantee $C^1$ global continuity.

• Orthogonal-Projection (OP) Method: Efficient when BC segments reside in hyperplanes, this approach uses signed distance functions $\bar{\phi}_i(x)$ and orthogonal projections $\mathcal{N}(x;\bar{\phi}_i)$ onto boundary planes. The network trunk evaluates $\Psi_i$ and $f_i$ on the boundary hyperplane and reconstructs the solution in the interior. This ensures exact boundary satisfaction with minimal structural overhead.

2.2 Kernel Correction in Integral Neural Operators

BOON techniques can act as plug-in correction modules for kernel-based neural operators as described in (Saad et al., 2022). The kernel $K$ in

$$u(x, t) = T[a](x) = \int_{\Omega} K(x,y) a(y) dy, \qquad (a(y) = u_0(y))$$

is modified so that the resulting operator $T$ enforces Dirichlet, Neumann, or periodic BCs at boundary points/discrete indices. This is achieved via low-rank updates to $K$'s rows/columns:

• For Dirichlet, $K_{\text{new}}(x_0,y)$ is replaced by $$\frac{\alpha_D(x_0,t)}{\alpha_D(x_0,0)} \delta(y-x_0)$$.
• For Neumann, $\partial_x K_{\text{new}}(x_0,y)$ is modified analogously.
• For periodic, the boundary rows are tied via $K_{\text{new}}(x_0,y) = K_{\text{new}}(x_{N-1},y)$.

These modifications are performed efficiently on the kernel matrix with $O(N)$ extra space and three additional $\mathcal{K}$ calls per layer, ensuring exact enforcement without additional trainable parameters.

2.3 Data-Driven Boundary Operator Learning

The MAD-BNO (Mathematical Artificial Data–Boundary Neural Operator) framework (Wu et al., 16 Jan 2026) constructs a linear map on the boundary:

$N_{\theta}: (g_D, h_N) \mapsto (h_D, g_N)$

where $(g_D, h_N)$ are Dirichlet and Neumann traces on $\partial \Omega_D$, $\partial \Omega_N$. The network consists of a single $400 \times 400$ weight matrix, and all training data pairs are synthesized via analytical evaluation of fundamental solutions. The interior solution is then reconstructed via classical boundary integral formulations, ensuring global PDE and BC consistency.

3. Algorithmic Workflows and Architectural Components

The primary BOON implementation pathways are summarized as follows:

Enforcement Mode	Algorithmic Steps & Components	Applicability
Solution-Structure (GLSS/OP)	Preprocess domain and BC segments; compute distance and normalized functions; implement structure layer as last step in neural operator; loss = PDE residual only.	Arbitrary piecewise $C^1$ boundaries (GLSS), hyperplanar BCs (OP)
Kernel Correction	Wrap each kernel-vector multiplication with BC-specific correction, using at most three calls per layer on discrete grids.	Any linear operator; BC location indexes required
MAD-BNO	Synthesize analytic BC data; train a bias-free linear layer on the boundary; reconstruct interiors by boundary integral.	PDEs with known fundamental solution (Laplace, Poisson, Helmholtz)

In all cases, BOON methods are hyperparameter-efficient (no additional tunables), modular, and preserve the existing neural operator backbone such as FNO or DeepONet.

4. Theoretical Properties: Regularity, Stability, Convergence

Rigorous analysis establishes:

• For solution-structure layers, the ansatz $u(x)$ satisfies all Dirichlet and Robin/Neumann BCs exactly for smooth $\Psi_i,\Psi_{\text{rem}}$; $C^1$-regularity on piecewise boundaries is ensured by weighted interpolation and corner matching (Göschel et al., 28 Oct 2025).
• No mesh-dependent penalty constants are required; loss landscape is as well-conditioned as classic collocation methods, a significant improvement over penalty-based BC enforcement.
• Kernel correction approaches (BOON wrappers) guarantee the existence of a corrected kernel enforcing boundary identities on the discretization grid, with bounded influence on the interior; uniqueness is restored for the learned operator (Saad et al., 2022).
• For MAD-BNO, boundary-to-boundary mapping is linear for elliptic PDEs, and interior error is controlled by the operator norm of the learned boundary map and the quadrature error in the boundary integral (Wu et al., 16 Jan 2026). The approach is extensible in principle to three-dimensional and complex domains.

5. Comparative Performance on Benchmark Tasks

Empirical experiments across distinct PDEs and geometries demonstrate substantial gains:

• Scalar Darcy Flow on L-shaped Domain (Göschel et al., 28 Oct 2025):
  • Operator training: OP L²-error $0.02 \pm 0.01$, GLSS $0.03 \pm 0.04$, semi-weak $0.03 \pm 0.03$, weak $0.05 \pm 0.05$.
  • Fine-tuned: OP/GLSS $\approx 0.01$, semi-weak $0.03$, weak $0.04$.
  • PINN-style: GLSS $0.02$, OP $0.04$, weak $0.13$.
• Navier–Stokes (Re ≈ 100) around Cylinder (Göschel et al., 28 Oct 2025):
  • After $4000$ epochs: $u$/$v$/$p$-errors $\approx 0.01$ (GLSS/OP), $0.05$ (semi-weak), $0.08$ (weak).
  • Drag/pressure-drop error within $1\%$ for GLSS/OP, $16\textrm{–}42\%$ for weak/semi-weak.
• BOON kernel correction (FNO, Burgers’, Navier-Stokes, Heat Equation) (Saad et al., 2022):
  • Burgers’ equation: BOON $0.000084$ (boundary error $=0$) vs FNO $0.0028$.
  • Stokes’ problem: BOON $0.0089$ vs FNO up to $0.0273$ (boundary error up to $0.0135$).
  • Improvement factors of $2\times$–$30\times$ in relative $L^2$ error with zero boundary norm.
• MAD-BNO (Laplace, Poisson, Helmholtz, 2D and 3D) (Wu et al., 16 Jan 2026):
  • Boundary-to-boundary: Neumann error $1.4 \times 10^{-2}$–$2.9 \times 10^{-2}$.
  • Interior via integral: $3 \times 10^{-3}$–$5 \times 10^{-2}$.
  • Training time reduction: Dirichlet Laplace 2.61 h (BOON) vs. 14.93 h (MAD-DeepONet), 31.09 h (PI-DeepONet); similar or better accuracy.

All BOON methods result in a $\lesssim30\%$ computational overhead (for structure layers) or negligible additional cost (kernel correction), with dramatically better boundary fidelity and faster error decay relative to weak or semi-weak approaches.

6. Guidance for Method Selection and Limitations

• Weak BC enforcement (penalty terms): simplest implementation, but incurs loss of precision and slower convergence—unsuitable if strict boundary satisfaction is essential.
• Semi-weak enforcement (exact Dirichlet + penalty Robin): low incremental complexity, appropriate for uncertain Robin data.
• Strong enforcement:
  • GLSS: For domains with arbitrary piecewise-$C^1$ geometry and mixed BCs; complexity scales with segment count.
  • OP: For domains with straight BC segments; reduced parameter count, fastest inference.
  • Kernel correction: Universally compatible with FNO, MGNO, multi-step, and temporal operators; no extra tunable weights needed.
  • MAD-BNO: Optimal where boundary integral representations exist and the fundamental solution is known.

A plausible implication is that BOON frameworks are extensible to nonlinear or multi-physics problems, provided solution-structure or kernel-correction analogues can be constructed; extension to highly irregular (e.g., fractal) boundaries or complex coupled PDE systems may require further methodological generalization.

7. Related Methodologies and Extensions

BOON intersects with hybrid physics-informed/deep learning paradigms, e.g., PINNs, PINOs, APINO, and boundary integral networks. It connects directly to R-function methods, transfinite interpolation, and domain decomposition. The use of synthetic (MAD) data in BOON (MAD-BNO) highlights the trend toward synthesizing training data via analytic or physical priors, circumventing the need for explicit PDE solves in training.

The approach in (Wu et al., 16 Jan 2026) demonstrates that pure boundary-based operator learning, combined with integral recovery, achieves parity or better performance compared to full-domain neural operators with significant reductions in training cost. The demonstrable extension to 3D Helmholtz equations with complex-valued boundary data further establishes BOON as a generalizable paradigm for operator learning under explicit boundary constraints.

---

References:

• "Enforcing boundary conditions for physics-informed neural operators" (Göschel et al., 28 Oct 2025)
• "Guiding continuous operator learning through Physics-based boundary constraints" (Saad et al., 2022)
• "Operator learning on domain boundary through combining fundamental solution-based artificial data and boundary integral techniques" (Wu et al., 16 Jan 2026)