NeuralOperator 2.0.0: Discretization-Agnostic Learning

• NeuralOperator 2.0.0 is a modular, discretization-agnostic framework for learning operators that map between continuous function spaces.
• It integrates diverse architectures including Fourier, graph, and geometry-informed operators to address nonlocality and ensure discretization convergence.
• The framework supports physics-informed, data-free training and probabilistic inference, enabling efficient PDE surrogates, optimal control, and finite-dimensional interpolation.

Searching arXiv for papers on NeuralOperator and related neural-operator framework papers. NeuralOperator 2.0.0 denotes the modern operator-learning stack centered on NeuralOperator, an open-source Python library for learning mappings between function spaces rather than finite-dimensional Euclidean spaces. In this formulation, neural operators are trained and inferenced on input and output functions given at various discretizations, with discretization convergence as a defining property. Recent work treats this stack not simply as a repository of Fourier Neural Operators or DeepONets, but as a modular framework for discretization-agnostic architecture design, geometry-aware kernels, physics-informed and data-free training, probabilistic inference, and task-adaptive operator learning (Kossaifi et al., 2024, Berner et al., 12 Jun 2025).

1. Conceptual foundations and software scope

NeuralOperator is built on top of PyTorch and provides core neural operator architectures, building-block layers, datasets and data loaders tailored to operator-learning benchmarks, and trainer and DataProcessor modules for training, inference, normalization, and discretization handling. The library also exposes efficiency tools such as tensorization, mixed precision, multi-grid domain decomposition, and incremental learning (Kossaifi et al., 2024). In the operator-learning setting, the central object is a parametric map

$\mathcal{G}_\theta : U \to V,$

where $U$ and $V$ are Banach or Sobolev spaces of functions on a domain $D$, rather than $\mathbb{R}^n\to\mathbb{R}^m$ maps (Le et al., 2024).

A generic neural-operator template uses a lifting operator $P$, operator-valued hidden layers, and a projection $Q$: $$v_0(x)=P(u)(x),\qquad v_{i+1}(x)=W_i v_i(x)+\sigma\!\left(\int_D \kappa_\theta(x,y)\,v_i(y)\,dy\right),\qquad \mathcal{G}_\theta(u)(x)=Q(v_L(x)).$$ This abstraction encompasses FNO-style spectral operators, learned integral kernels, and graph-based message-passing analogues (Le et al., 2024). A complementary design recipe for “extending neural architectures to function spaces” replaces discrete weight matrices by continuous kernels, introduces coordinates explicitly, fixes receptive fields in domain coordinates rather than index space, and discretizes integrals with quadrature weights. This formulation is intended to preserve discretization-agnosticity and discretization convergence under mesh refinement (Berner et al., 12 Jun 2025).

Within this view, NeuralOperator 2.0.0 is best understood as both a software package and an architectural doctrine: functions are first-class inputs and outputs, kernels are continuous objects evaluated on discretizations, and operator layers are expected to remain meaningful as grids, meshes, or point clouds change.

2. Architectural families and internal representations

The library-level core includes Graph Neural Operators, Fourier Neural Operators, Tensorized Fourier Neural Operators, Spherical Fourier Neural Operators, Geometry-informed Neural Operators, U-shaped architectures, and localized kernel layers (Kossaifi et al., 2024). These families differ primarily in how they parameterize the nonlocal operator $\mathcal{K}_\theta$: FNO uses Fourier multipliers on regular grids, SFNO replaces planar FFTs with spherical harmonic transforms, GNO and GINO use graph or geometry-aware kernels on irregular domains, and TFNO compresses spectral kernels through tensor decompositions (Kossaifi et al., 2024).

The theoretical literature has progressively clarified that the decisive ingredients are nonlocality and nonlinearity. The “averaging neural operator” shows that even a minimal nonlocal mechanism—a spatial average—combined with nonlinear lifting and projection suffices for universal approximation of continuous operators on compact sets. This result reframes FNO: many Fourier modes are not required for universality in principle, because the essential requirement is a nonlocal feature channel, not an arbitrarily rich spectral expansion (Lanthaler et al., 2023).

Recent architectural work extends this base in several directions. DPNO organizes a basic operator block in parallel two paths which are similar with ResNet and DenseNet, rather than purely stacking blocks. On Burgers’ equation, Darcy Flow Equation and the 2d Navier-Stokes equation, it reports relative improvement of over 30% on certain standard test cases and applies successfully to both DeepONet and FNO (Wang et al., 17 Jul 2025). HNO, by contrast, injects analytic-signal structure: it first maps the input to a Hilbert-transformed representation, then applies spectral convolution in that domain, with the stated aim of making instantaneous amplitude and phase explicit inductive biases for causal, phase-sensitive, and non-stationary systems (Pordanesh et al., 6 Aug 2025).

A common simplification equates neural operators with FNO on rectangular grids. The current architecture landscape is broader: it includes branch–trunk decompositions, kernel-integral layers, graph formulations, manifold spectral bases, tensorized kernels, and hybrid local–global stacks.

3. Training paradigms: supervised, physics-informed, and data-free

The classical training mode is supervised operator regression on paired input–output functions. Physics-informed operator learning generalizes PINNs from single PDE instances to operator families. A standard hybrid objective is

$$\mathcal{L}=\mathcal{L}_{\text{data}}+\mathcal{L}_{\text{physics}},$$

with $U$0, where the residual or variational form of the governing PDE is evaluated on the operator output $U$1 (Goswami et al., 2022). The same literature emphasizes that physics-informed operator learning need not be restricted to pointwise residual penalties: variational energies are used for brittle fracture and Darcy flow, and multiple independently pre-trained DeepONets can be coupled as components of multi-physics systems (Goswami et al., 2022).

A major development is Walk-on-Spheres Neural Operator (WoS-NO), a data-free, physics-informed training backend for neural operators on linear elliptic PDE families. Instead of relying on expensive FEM datasets or higher-order automatic differentiation of PDE residuals, WoS-NO uses online Monte Carlo physics: during training, the PDE solution is estimated via Walk-on-Spheres at random spatial points and used as noisy labels. The operator is trained against these weak supervisions with an $U$2 loss, with no PDE residuals or derivatives in the loss (Viswanath et al., 1 Mar 2026). For the same number of training steps, the reported improvements are up to $U$3 in $U$4-error compared to standard physics-informed training schemes, up to $U$5 in training speed, and reductions of up to $U$6 in GPU memory consumption (Viswanath et al., 1 Mar 2026).

The quantitative tables in WoS-NO make the tradeoff concrete. On a 2D Poisson family trained for 20k steps, WoS-NO reports time $U$7 min, MSE $U$8, and GPU memory $U$9 MB, compared with PINO at $V$0 min, $V$1, and $V$2 MB (Viswanath et al., 1 Mar 2026). On a 3D variable-coefficient Poisson family at 50k steps, WoS-NO reports $V$3 min, $V$4, and $V$5 MB, compared with PINO at $V$6 min and $V$7 MB (Viswanath et al., 1 Mar 2026).

This broadens the meaning of “physics-informed” in NeuralOperator 2.0.0. Physics-informed training can be residual-based, variational, Monte Carlo weakly supervised, or mixed. The unifying idea is operator learning over PDE families, not commitment to a single loss construction.

4. Discretization, locality, manifolds, and geometric generalization

Discretization-independence is the defining software and mathematical property of the stack. The library description stresses that a neural operator, with a fixed set of parameters, can be applied to input functions given at any discretization, and that outputs differ only by a discretization error that converges to zero as the discretization is refined (Kossaifi et al., 2024). The practical recipe behind this is explicit in architecture-conversion work: kernels act on coordinates, quadrature weights approximate integrals, and receptive fields are defined in physical coordinates rather than index offsets (Berner et al., 12 Jun 2025).

Recent research has sharpened the role of locality. “Neural Operators with Localized Integral and Differential Kernels” proves that, under appropriate scaling, CNN kernels converge to differential operators, while DISCO-style compactly supported kernels define local integral operators that retain operator-learning properties and therefore resolution generalization. Adding these layers to FNOs reduces relative $V$8-error by $V$9–$D$0 on turbulent 2D Navier-Stokes and the spherical shallow water equations (Liu-Schiaffini et al., 2024). This directly addresses the over-smoothing associated with purely global spectral layers.

On non-Euclidean domains, NORM generalizes neural operators to Riemannian manifolds by replacing Fourier modes with Laplace–Beltrami eigenfunctions. The core encoder–approximator–decoder block $D$1 has universal approximation on compact subsets of $D$2 function spaces on manifolds, and the architecture is discretization-independent in the sense that parameters live in spectral coefficient space rather than mesh-node space (Chen et al., 2023). Empirically, NORM is evaluated on Darcy flow on irregular 2D domains, pipe turbulence, 2D boundary $D$3 3D heat transfer, composite workpiece deformation, and blood flow in the human aorta, consistently emphasizing manifold-native operator learning (Chen et al., 2023).

A further step is the kernel-integral view of variable geometry. “Geometric Generalization of Neural Operators from Kernel Integral Perspective” recasts operator learning on variable geometries as the approximation of geometry-dependent kernel operators, potentially with singularities, and proposes a multiscale point-cloud neural operator inspired by Ewald summation. The architecture splits long-range Fourier structure from short-range geometric corrections involving normals, tangential derivatives, and curvature, and numerical experiments demonstrate robust generalization across diverse geometries and in a large-scale three-dimensional fluid dynamics example (Han et al., 2 Feb 2026).

Taken together, these developments show that NeuralOperator 2.0.0 is not confined to Cartesian FFTs. It includes localized differential operators, compactly supported integral kernels, Laplace–Beltrami spectral bases, and point-cloud kernel integrals as principled mechanisms for geometry and resolution transfer.

5. Probabilistic inference and finite-dimensional interpolation

A decisive extension beyond deterministic operator regression is FunDPS, a function-space diffusion model with plug-and-play guidance for PDE inverse problems. FunDPS trains an unconditional discretization-agnostic denoising model using neural operator architectures and conditions at inference through a gradient-based guidance mechanism derived from an extension of Tweedie’s formula to infinite-dimensional Hilbert spaces (Yao et al., 22 May 2025). Across five PDE tasks with only $D$4 observation, it reports an average $D$5 accuracy improvement over state-of-the-art fixed-resolution diffusion baselines while reducing sampling steps by $D$6 (Yao et al., 22 May 2025). With ReNoise and multi-resolution inference, the reported wall-clock speedup over DiffusionPDE is approximately $D$7 (Yao et al., 22 May 2025).

This probabilistic turn changes the role of a neural-operator library. Neural operators become denoisers and priors on function spaces, not only regressors of mean solutions. Posterior sampling, uncertainty quantification, sparse-observation reconstruction, and inverse problems become native use cases rather than external wrappers.

An orthogonal reframing appears in “Neural Operators as Efficient Function Interpolators”. By introducing an auxiliary base-space, any finite-dimensional function can be viewed as an operator acting by composition on functions of that base-space, allowing TFNOs to serve as parameter-efficient interpolators for ordinary finite-dimensional regression tasks (Niarchos et al., 8 May 2026). On structured scientific data, a two-dimensional TFNO ensemble on the nuclear chart reaches a held-out root-mean-square error of $D$8 keV for residual correction to the WS4 mass model, with each model using approximately $D$9 parameters (Niarchos et al., 8 May 2026). This work broadens the remit of the stack from PDE surrogates to a scalable framework for finite-dimensional function interpolation.

6. Control, scientific applications, and broader ecosystem functions

Neural operators are no longer restricted to coefficient-to-solution maps. In multi-task control, the learned operator maps task-defining functions—costs, dynamics, or obstacles—to feedback policies. “Neural Operators for Multi-Task Control and Adaptation” uses a permutation-invariant branch–trunk architecture in which the branch encodes a set of task samples and the trunk evaluates state–time queries, thereby approximating solution operators for optimal control problems (SeWell et al., 3 Apr 2026). A single operator trained via behavioral cloning generalizes to unseen tasks, out-of-distribution settings, and varying amounts of task observations, while branch-only, last-layer, full-network, and meta-trained adaptation strategies provide different data–compute tradeoffs (SeWell et al., 3 Apr 2026).

The application range represented in the current ecosystem is correspondingly broad. The library paper points to plasma surrogate modeling using FNOs, real-time high-resolution CO$\mathbb{R}^n\to\mathbb{R}^m$0 storage prediction, FourCastNet-style global weather modeling, and huge ensembles for weather forecasts using SFNOs (Kossaifi et al., 2024). Manifold and geometry-aware work adds irregular-domain Darcy flow, heat transfer from 2D boundaries to 3D solids, and patient-specific aortic blood flow (Chen et al., 2023). Probabilistic operator learning targets PDE inverse problems under severe data scarcity (Yao et al., 22 May 2025). Function-interpolation work extends the paradigm to analytic benchmarks and nuclear masses (Niarchos et al., 8 May 2026).

This suggests that “NeuralOperator 2.0.0” is best viewed as an ecosystem function: the same operator-learning substrate supports deterministic PDE surrogates, data-free weak supervision, posterior sampling, control-law approximation, and interpolation on structured scientific domains.

7. Limitations, misconceptions, and open problems

Several recurrent limitations remain. Theoretical analyses often assume continuity of the target operator, compact subsets of Sobolev or Banach spaces, Lipschitz components, and, in many results, regularity $\mathbb{R}^n\to\mathbb{R}^m$1 so that Sobolev embeddings can control nonlinearities. Non-smooth PDE solutions, shocks, and discontinuities are therefore not fully captured by the dominant Sobolev-space theory (Le et al., 2024). Finite-capacity lower bounds also make clear that universality does not imply arbitrarily small error at fixed depth and width (Le et al., 2024).

Method-specific restrictions are equally important. WoS-NO is experimentally validated only on linear elliptic Poisson-type problems, handles Dirichlet boundary conditions naturally, and relies on amortization to control the high variance of low-sample Monte Carlo supervision; early termination and caching introduce bias (Viswanath et al., 1 Mar 2026). NORM assumes access to Laplace–Beltrami eigenfunctions on compact manifold-like geometries, and extending the method to families of varying geometries while sharing a global operator remains an open direction (Chen et al., 2023). FunDPS requires guidance-weight tuning, and extension to more complex domains, unstructured meshes, and 3D settings is still to be fully explored (Yao et al., 22 May 2025).

The literature also corrects several common misconceptions. Neural operators are not confined to regular grids or Fourier layers; graph, manifold, localized-kernel, point-cloud, and branch–trunk formulations are all standard within the current stack (Kossaifi et al., 2024, Chen et al., 2023). Physics-informed operator learning is not synonymous with higher-order residual autodiff; variational losses and derivative-free weak supervision are already established alternatives (Goswami et al., 2022, Viswanath et al., 1 Mar 2026). Nor is the field purely deterministic: function-space diffusion models and posterior samplers now occupy a central place in the ecosystem (Yao et al., 22 May 2025).

Open problems follow directly from these tensions. Current theory still needs sharper dependence on operator norms, grid resolution, and function regularity; geometry-varying operator learning requires stronger guarantees beyond kernel-level approximation; and many promising directions—non-smooth function spaces, adaptive multiscale geometry, nonlinear PDEs on variable domains, and architecture-level contraction or energy-descent guarantees—remain active research topics (Le et al., 2024, Han et al., 2 Feb 2026).

In that sense, NeuralOperator 2.0.0 is less a fixed endpoint than a consolidation phase: a mature software base for operator learning, coupled to a rapidly expanding research program that extends operator methods from spectral PDE surrogates to geometry-aware computation, physics-informed and data-free training, probabilistic inference, adaptive control, and structured scientific interpolation.