Neural-Operator Element Method (NOEM)

• Neural-Operator Element Method (NOEM) is a hybrid approach that blends neural network learning with classical variational discretizations to solve PDEs.
• It replaces or augments traditional basis functions with data-driven neural operators, achieving significant acceleration and maintaining robust accuracy.
• NOEM frameworks, including augmented FE methods, mesh-based predictors, and trainable basis functions, enable efficient solutions for parametric, nonlinear, and multiscale problems.

The Neural-Operator Element Method (NOEM) is a class of hybrid numerical schemes that integrate neural operator learning with finite element or mesh-based variational discretizations for the solution of partial differential equations (PDEs). These approaches seek to synergize the local accuracy, adaptivity, and convergence theory of classical meshed methods (FEM, VEM, DG) with the high-dimensional function-approximation capacity, mesh invariance, and parametric learning efficiency of neural operators. The NOEM paradigm encompasses both data-driven surrogate acceleration of established solvers and new trial/test spaces in which neural-network-based elements replace or augment classical basis functions, projections, and stabilization concepts. Empirical and theoretical studies demonstrate that NOEM can provide orders-of-magnitude acceleration over traditional methods without compromising accuracy or stability, especially for parametric, nonlinear, and multiscale PDEs.

1. Fundamental Principles and Methodological Frameworks

NOEM is rooted in the reinterpretation of classical element-based Galerkin schemes—FEM, VEM, and DG—in the language of operator learning. The distinguishing feature is the hybridization of explicit mesh-based trial spaces with neural operators trained to either parametrize local solution maps or directly produce basis representations. Several distinct NOEM flavors have been proposed:

• Neural-Operator Augmented FE Methods: The FE mesh is used as a backbone; within selected subdomains, standard elements are replaced by neural-operator elements (NOEs) that map boundary values to local interior solutions via pre-trained neural operators. The trial space is a direct sum of standard FE bases and NOE-based surrogates. Variational assembly leverages both explicit polynomials and neural predictions, with global degrees of freedom including both nodal values and NOE sensor points (Ouyang et al., 23 Jun 2025).
• Mesh-Based Neural Coefficient Predictors: Neural networks are trained to map problem data (forcing, coefficients) to the expansion coefficients of a fixed FE/DG basis. The loss is the residual of the (parametric) algebraic weak form, requiring no paired solution data (Hong et al., 2024, Chawla et al., 7 Jan 2026).
• Trainable Finite-Element-Like Bases: The representation network constructs explicit local basis functions (e.g., piecewise-linear “hats” or polygonal/harmonic VEM bases) within shallow networks. Basis function locations and supports are made learnable and data-adaptive, yielding effective parameter compression and local mesh refinement features (Zhang et al., 30 Oct 2025, Wang et al., 23 Apr 2025, Berrone et al., 8 Jul 2025, Berrone et al., 2023).
• Hybrid Iterative Schemes: Mesh-free neural operators first provide a coarse solution prediction, which is then corrected and refined by a few steps of a classical mesh-based iterative solver (e.g., Newton or conjugate-gradient). Neural outputs serve as effective "warm starts" and accelerate nonlinear or parametric solves (Wang et al., 6 Jan 2026, Taghikhani et al., 10 Nov 2025, He et al., 2023).

These frameworks preserve variational consistency and leverage the existing theory and infrastructure of established discretizations while extending their applicability using modern machine learning.

2. Variational and Algebraic Formulation

NOEM methods retain the essential structure of weak (Galerkin) formulations. For model problems such as

$$-\nabla\cdot(\kappa(x)\nabla u(x)) = f(x)\ \text{in } \Omega, \quad u|_{\partial\Omega}=g,$$

the weak form seeks $u\in V$ such that

$$a(u,v) = \int_\Omega \kappa \nabla u \cdot \nabla v = L(v) = \int_\Omega f v,\quad \forall v\in V_0.$$

In NOEM, the trial/test space $V_h^{\text{NOEM}}$ may consist of standard FE shape functions on $\Omega^{\text{FE}}$ and neural-operator elements on $\Omega^{\text{NOE}}$, either as direct surrogates of local PDE maps or as mesh-based functions generated by neural nets. The global discrete problem determines the combined coefficient vector (standard DOFs and NOE sensor/boundary data) by minimizing a discrete energy functional or solving the assembled system via Newton's method (Ouyang et al., 23 Jun 2025).

In coefficient-predictor architectures, the neural net maps data $\omega$ to the coefficient vector $\alpha(\omega)$, and the predicted field is

$$\hat u_h(x;\omega) = \sum_{k=1}^{N_h} \hat\alpha_k(\omega)\,\phi_k^h(x),$$

with the loss functional derived from the discrete variational residual: $$L(\alpha) = \mathbb{E}_\omega\|A\alpha(\omega) - F(\omega)\|^2_{\ell^2}$$ with $A$ and $F(\omega)$ computed using the FE mesh (Hong et al., 2024, Chawla et al., 7 Jan 2026).

For elementwise-learned basis approaches, the neural net predicts basis coefficients for local trial functions, with the full system assembled via explicit Gaussian quadrature and standard FEM/VEM logic (Zhang et al., 30 Oct 2025, Berrone et al., 2023, Berrone et al., 8 Jul 2025).

3. Neural Operator Architectures and Training Strategies

The neural operators and relevant architectures differ according to the NOEM variant:

• Branch/Trunk Networks: DeepONet-type structures with branch nets absorbing parametric/boundary data and trunk nets handling spatial queries, typically with ReLU/tanh activations and depths of $3$-$7$ layers, widths $50$-$300$ neurons. Training is supervised or physics-informed depending on the variant (Ouyang et al., 23 Jun 2025, He et al., 2023).
• Mesh-wise Coefficient Predictors: Fully-connected MLPs or shallow CNNs, inputting discretized field quantities (material, loading, BCs) and outputting vectors of coefficients for the nodal/elemental expansion (Hong et al., 2024, Chawla et al., 7 Jan 2026).
• Finite Element Representation Networks (FERN): Shallow networks define explicit piecewise-linear “hat” functions whose centers and widths are trainable. Only coefficients and geometric parameters are learned; the basis is fully encoded as a compact ReLU network, guaranteeing exact finite-element assembly and supporting adaptation to localized solution features (Zhang et al., 30 Oct 2025).
• Neural-Virtual-Element Approaches: For polygonal meshes/VEM, networks are trained to directly map local geometry encodings (vertex positions) to coefficients in a harmonic/polynomial basis, thus realizing element-wise closed-form basis functions without recourse to projection or stabilization operators (Berrone et al., 8 Jul 2025, Berrone et al., 2023).

Training can be data-driven (where reference solutions are available), physics-constrained (using PDE residual or weak form as the loss, with no solution labels required), or hybrid (when both physics and data are blended). Offline training is typically performed per element type or per geometry class, and in the case of coefficient or basis function prediction, generalization to unseen parametric regimes and mesh discretizations is a core design feature.

4. Theoretical Error and Convergence Analysis

Recent theoretical work establishes detailed a priori and generalization error bounds for mesh-based neural operator methods. For linear elliptic PDEs, suppose $A$ is the FE stiffness matrix and $\kappa(A)$ its condition number. Then, for network capacity $n$ and $M$ training samples, the composite error satisfies (Hong et al., 2024): $$\mathbb{E}[\|u-u_{h,n,M}\|_{L^2}] \lesssim h^{\ell+1} + \frac{\kappa(A)^{1+d}}{\sqrt{n}} + \frac{\kappa(A)^{1+3d/2}}{\sqrt{M}}.$$ The mesh error decays with $h$, while neural approximation and generalization terms are controlled by network size and sample count, amplified by $\kappa(A)$ (which itself scales like $h^{-2}$). Preconditioning with a sparse approximate inverse drastically reduces this sensitivity. In the DG-FEONet framework, convergence of the elementwise neural coefficient predictor to the true DG solution is established under standard network capacity arguments, with empirical generalization tracked by Rademacher complexity (Chawla et al., 7 Jan 2026).

For locally adaptive trainable-basis methods (FERN), performance matches or exceeds global operator learners in L$^2$ accuracy and variance, with $3$-$10\times$ fewer parameters and the ability to adaptively refine basis support near solution features (Zhang et al., 30 Oct 2025).

5. Numerical Experiments and Comparative Performance

NOEM approaches have been extensively tested on prototypical and challenging PDE classes encompassing:

• Elliptic and parabolic model problems: 1D and 2D Poisson, advection-diffusion-reaction, multiscale elliptic, and nonlinear p-Laplace.
• Solid and continuum mechanics: Linear elasticity, hyperelasticity, multiscale and phase-contrast microstructures, stress analysis on complex geometries (Wang et al., 6 Jan 2026, Taghikhani et al., 10 Nov 2025, Ouyang et al., 23 Jun 2025, Berrone et al., 8 Jul 2025).
• Parametric and non-smooth/irregular problems: Discontinuous coefficient fields, sharp gradient localization (shocks), inclusions, and highly heterogeneous domains (Chawla et al., 7 Jan 2026, Berrone et al., 2023).

Empirical findings include:

• Sub-$1$–$2\%$ relative errors for local NOEM surrogates on complex geometry problems, with up to $10\times$ reduction in DOFs and an order-of-magnitude acceleration over full FE solves (Ouyang et al., 23 Jun 2025, He et al., 2023).
• Robust capture of discontinuities and parameterized generalization across wide families of PDEs (DG-FEONet, FERN).
• In hybrid iterative (Newton-initialized) frameworks, $1$–$3$ Newton corrections suffice to reach full FE accuracy, yielding up to $8\times$ wall-clock speedup in large 3D nonlinear mechanics problems (Taghikhani et al., 10 Nov 2025).
• For operator-learned polygonal VEM, NOEM variants eliminate stabilization artifacts and outperform standard VEM in both convergence constant and robustness, especially for nonlinear or highly distorted meshes (Berrone et al., 8 Jul 2025, Berrone et al., 2023).

A summary of representative timing and error statistics is provided in tabular format in the referenced works (e.g., (Wang et al., 6 Jan 2026, Ouyang et al., 23 Jun 2025, Zhang et al., 30 Oct 2025)).

6. Extensions, Applications, and Limitations

NOEM is directly extensible to a variety of application domains and methodological enhancements:

• Parametric and Uncertainty Quantification: High-dimensional parameter fields (e.g., multiscale material properties, random media) can be handled by training neural operator surrogates over large ensembles, enabling accelerated Monte Carlo or Bayesian studies (He et al., 2023, Wang et al., 6 Jan 2026).
• Multiscale, Multiphysics, and Operator Foundation Modeling: NOEM is applied to multiscale PDEs by constructing reusable NO elements for complex subdomains; time-dependent and coupled-physics extensions are under active development (Ouyang et al., 23 Jun 2025, Wang et al., 6 Jan 2026).
• Hybrid Data-Physics Training: PINN-style regularization, physics-informed losses, and self-adaptive target-weighting can be incorporated to achieve mesh/geometry/generalization invariance (Wang et al., 6 Jan 2026, Zhang et al., 30 Oct 2025).
• Adaptive Mesh and Basis Construction: FERN and related NOEMs decode data-driven mesh refinement behaviors; basis function widths and locations adapt to evolving solution features (shocks, fronts, inclusions) (Zhang et al., 30 Oct 2025).

Reported limitations include the need for problem-class-specific training, offline computational cost for neural operator construction (offset by subsequent reusability), absence of complete a priori error estimates in the nonlinear regime, and challenges in guaranteeing cross-element $C^0$ or $C^1$ regularity without specialized coupling strategies (Ouyang et al., 23 Jun 2025).

7. Relation to Broader Operator Learning and Numerical Analysis

NOEM sits at the intersection of several current trends:

• Operator Learning: It generalizes beyond pure mesh-free neural operators (DeepONet, FNO), embedding mesh structure, local PDE geometry, and FE-compliant enforcement of boundary and variational constraints (Wang et al., 23 Apr 2025, Ouyang et al., 23 Jun 2025).
• Mesh-Based Surrogates: FEONet and DG-FEONet bridge unsupervised physics-informed learning with the algebraic structure of classical methods, combining data efficiency with convergence theory (Hong et al., 2024, Chawla et al., 7 Jan 2026).
• Polygonal and Virtual Element Methods: Neural-approximated VEM eliminates projection and stabilization issues while remaining fully variational, leveraging geometric encoding in neural parameterizations (Berrone et al., 8 Jul 2025, Berrone et al., 2023).

A plausible implication is that NOEM, as an umbrella for these approaches, is positioned to become a foundational methodology enabling scalable, accurate, and reusable surrogates for large-scale scientific computing, design, and real-time multiphysics digital twins. The field remains rapidly evolving, with anticipated theoretical advances and increasing adoption in industry-scale simulation workflows.