Physics-Informed Neural Operator Framework

• Physics-Informed Neural Operator Framework is a class of deep learning methods that learn solution operators by integrating data with governing PDE, boundary, and initial condition constraints.
• It employs a multi-resolution training strategy that couples coarse data supervision with fine-grid physics enforcement to improve high-frequency accuracy and enable zero-shot super-resolution.
• The framework spans diverse architectures—such as Fourier, hypernetwork-enhanced, and wavelet-based operators—making it broadly applicable in turbulence, multiscale mechanics, and wave propagation.

Searching arXiv for recent and foundational papers on physics-informed neural operators. Search query: "physics-informed neural operator PINO" Physics-informed neural operator framework denotes a class of operator-learning methods for parametric PDEs in which the target is not a single solution field but a solution operator, while the training objective embeds governing equations, boundary conditions, and initial conditions. In the foundational PINO formulation, the operator is learned with a hybrid of data and physics, including PDE constraints imposed at a higher resolution than the available training data, and the same framework can also operate with no training data and only PDE constraints (Li et al., 2021). Subsequent work broadened the framework beyond Fourier backbones to layered hypernetworks, DeepONet-type operators, wavelet and convolutional operators, low-rank point-cloud operators, and variational or FE-based formulations, while retaining the defining principle that operator learning and physics enforcement are coupled in a single model class (Wang et al., 21 Jun 2025, Eshaghi et al., 2024, Schaffer et al., 9 Sep 2025).

1. Conceptual scope and relation to neighboring paradigms

Operator learning seeks a map between function spaces. In the standard parametric setting, one learns a solution operator

$$\mathcal{G}:\mathcal{A}\to\mathcal{U},\qquad a\mapsto u=\mathcal{G}(a),$$

or, for dynamical systems,

$$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$

so that new coefficients, forcings, or initial conditions can be queried without retraining (Li et al., 2021). Physics-informed neural operators inherit this operator-level objective but replace, augment, or regularize supervised regression with PDE residuals and BC/IC constraints.

This separates the framework from several adjacent methods. PINNs optimize a single instance-wise map $u_\theta(x,t)$ for one PDE instance and do not learn a reusable operator over parameter space; they are therefore inefficient for many-query settings and are reported to struggle on multi-scale dynamics because of optimization pathologies such as spectral bias and difficulty propagating BC/IC information (Li et al., 2021). Data-driven FNOs learn operators via spectral convolutions but do not explicitly enforce PDE physics unless combined with physics losses (Wang et al., 21 Jun 2025). DeepONet learns an operator through a branch network for coefficients and a trunk network for basis functions, but in the PI-DeepONet form it becomes part of the broader PINO family (Howard et al., 2023). Hyper-PINNs dynamically map parameters to network weights, yet the monolithic parameter-to-weight mapping can be computationally heavy and unstable, which motivated structured variants such as layered hypernetworks (Wang et al., 21 Jun 2025).

A common misconception is that PINO is synonymous with “FNO plus residual loss.” The term is used more broadly for operator-learning models that incorporate physics constraints, including Fourier, DeepONet, wavelet, convolutional, low-rank, and variational formulations (N et al., 2023, Eshaghi et al., 2024).

2. Mathematical structure of the framework

A generic parametric PDE can be written as

$$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$

with boundary and initial operators

$$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$

The operator-learning objective is to construct

$$\mathcal{G}_\theta:(\mu,f)\mapsto u_\theta(\cdot;\mu,f),$$

so that $u_\theta$ satisfies the PDE family across the parameter space (Wang et al., 21 Jun 2025).

The standard composite physics-informed loss is

$$\mathcal{L} = \lambda_{\mathrm{pde}}\mathcal{L}_{\mathrm{pde}} + \lambda_{\mathrm{bc}}\mathcal{L}_{\mathrm{bc}} + \lambda_{\mathrm{ic}}\mathcal{L}_{\mathrm{ic}} + \lambda_{\mathrm{data}}\mathcal{L}_{\mathrm{data}},$$

with

$$\mathcal{L}_{\mathrm{pde}} = \mathbb{E}_\mu\, \mathbb{E}_{(x,t)\in\Omega\times[0,T]} \big[\|R(x,t;\mu)\|^2\big],$$

and analogous BC, IC, and optional supervised terms (Wang et al., 21 Jun 2025). For example, in Burgers,

$$u_t+u\,u_x-\nu u_{xx}=0,\qquad R=u_t+u\,u_x-\nu u_{xx}.$$

In the original PINO formulation, the loss is explicitly multi-resolution: $$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$0 where $$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$1 is the coarse data grid and $$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$2 is the fine physics grid (Li et al., 2021). This coupling of coarse supervision and fine-grid residuals underlies zero-shot super-resolution and high-frequency correction.

Residual evaluation is architecture-dependent. FNO-based PINO often uses pseudo-spectral differentiation in periodic directions and finite differences in time; fully connected or DeepONet-style PINO commonly relies on autograd; wavelet and convolutional variants may replace AD with stochastic projection or finite-difference stencils (Li et al., 2021, N et al., 2023, Ma et al., 22 Jul 2025).

3. Architectural families

The framework has diversified primarily along the choice of operator parameterization, basis, and physics-enforcement mechanism.

Framework	Core mechanism	Distinctive feature
PINO (Li et al., 2021)	FNO backbone with multi-resolution physics loss	coarse data, fine-grid PDE constraints
LFR-PINO (Wang et al., 21 Jun 2025)	layered hypernetworks + Fourier-reduced weights	per-layer parameter generation and weight-space spectral truncation
PI-WNO (N et al., 2023)	wavelet-domain operator learning	space-frequency localization and stochastic projection gradients
VINO (Eshaghi et al., 2024)	variational neural operator	energy minimization on discretized elements
EquiNO (Eivazi et al., 27 Mar 2025)	FE-OL with POD-reduced operators	divergence-free stress modes and periodic displacement modes
PICNO (Ma et al., 22 Jul 2025)	convolutional neural operator for Helmholtz scattering	discrete PDE residual with 8th-order FD Laplacian
PILNO (Schaffer et al., 9 Sep 2025)	low-rank kernel operator on point clouds	discretization-independent encoder–decoder

Foundational PINO uses spectral convolution layers of the form

$$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$3

with truncation to a fixed set of modes and a pointwise lift/projection around the operator stack (Li et al., 2021). This design exploits the universal approximation and discretization-convergent properties attributed to FNO.

LFR-PINO replaces a monolithic parameter-to-weight hypernetwork by layer-specific hypernetworks,

$$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$4

and further compresses each weight tensor by retaining only the first $$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$5 Fourier coefficients before inverse reconstruction (Wang et al., 21 Jun 2025). The intended effect is to reduce the dimensionality of parameter-to-weight mapping while preserving the low/mid-frequency structure emphasized by the frequency principle.

CPNO, although not required for the basic PINO definition, illustrates a distinct basis-transformed branch of the framework. It replaces unstable monomial expansions with a Chebyshev spectral basis, uses parameter-dependent modulation at each layer, and evaluates the resulting series stably through Clenshaw recurrence (Chen et al., 2 Feb 2026). The paper ties this to logarithmic Lebesgue-constant growth and improved conditioning.

PI-WNO moves the operator kernel from Fourier space to wavelet space,

$$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$6

and estimates derivatives through a stochastic projection-based local regression rather than backpropagating through wavelet transforms (N et al., 2023). VINO instead abandons strong-form residuals and minimizes an energy functional on a discretized mesh, using analytically computed element-level derivatives and integrals (Eshaghi et al., 2024). EquiNO embeds equilibrium and periodicity as hard constraints through divergence-free stress POD modes and periodic displacement POD modes in multiscale FE$$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$7 settings (Eivazi et al., 27 Mar 2025).

These variants suggest that “physics-informed neural operator” is better understood as a design principle than as a single architecture: the operator backbone, basis, and residual representation are all mutable, but the operator-learning objective and direct embedding of physical laws remain the invariant core.

4. Training paradigms and optimization strategies

The training regimes in this literature span supervised, semi-supervised, and fully unsupervised settings. Foundational PINO can learn with coarse data plus fine-grid physics, but it also succeeds in no-data settings where only PDE constraints are imposed (Li et al., 2021). LFR-PINO emphasizes pre-training a universal solver across diverse PDE families and parameter ranges without labeled solutions, then optionally fine-tuning for higher precision on new tasks (Wang et al., 21 Jun 2025).

A recurrent strategy is staged or curriculum training. In stacked multifidelity operator training, a low-fidelity PI-DeepONet or PINN is trained first, then its frozen output becomes the low-fidelity input for the next stage, with linear and nonlinear correlators and an $$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$8 penalty encouraging linear correlation when adequate (Howard et al., 2023). A distinct continuation mechanism appears in multi-stage curriculum training for PhIS-FNO, where boundary conditions are enforced first, interior residuals are introduced later, and Adam is re-initialized at each stage; the paper argues analytically and empirically that resetting $$\mathcal{G}:\mathcal{A}\to C((0,T];\mathcal{U}),\qquad a\mapsto u(\cdot,t),$$9 and $u_\theta(x,t)$0 restores effective learning rates and avoids gradient stagnation (Marcandelli et al., 2 Feb 2026).

Residual evaluation is another major axis of optimization design. Foundational PINO uses pseudo-spectral or FD schemes on fine grids and, for non-periodic problems, Fourier continuation (Li et al., 2021). PICNO computes the Helmholtz residual with an eighth-order finite-difference Laplacian, explicitly avoiding memory-intensive AD in a convolutional operator (Ma et al., 22 Jul 2025). PI-WNO uses stochastic projection gradients, while VINO removes high-order AD from the loss entirely by differentiating FE shape functions analytically (N et al., 2023, Eshaghi et al., 2024).

Hybrid supervision remains important when pure physics losses admit spurious or non-unique solutions. SPINONet reports that limited data supervision improves performance in challenging regimes where purely physics-informed training may converge to spurious solutions, and its Eikonal example uses sparse supervision to prevent flipped-sign or distorted solutions (Garg et al., 23 Mar 2026). A plausible implication is that unsupervised PINO is not intrinsically preferable to hybrid PINO; rather, the choice depends on PDE identifiability, boundary information, and optimization stiffness.

5. Applications and reported empirical behavior

Reported applications span canonical PDE benchmarks, turbulence, phase-field evolution, multiscale mechanics, Helmholtz wavefields, inverse scattering, and power-system dynamics.

Foundational PINO reported that in chaotic Kolmogorov flow, when trained on $u_\theta(x,t)$1 data and physics on $u_\theta(x,t)$2, data-only FNO obtained $u_\theta(x,t)$3 relative $u_\theta(x,t)$4 error across resolutions, whereas data+PDE loss obtained $u_\theta(x,t)$5–$u_\theta(x,t)$6; in Burgers, training on $u_\theta(x,t)$7 data and $u_\theta(x,t)$8 physics yielded $u_\theta(x,t)$9 relative $$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$0 error at $$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$1 with no degradation under zero-shot super-resolution (Li et al., 2021). LFR-PINO reported 22.8%–68.7% error reduction compared to baselines across anti-derivative, advection, Burgers, and diffusion–reaction, while frequency-domain reduction lowered memory by 28.6%–69.3% compared to Hyper-PINNs (Wang et al., 21 Jun 2025).

In multiscale mechanics, EquiNO reported speedup factors exceeding 8000-fold compared to traditional methods and stress errors typically below about $$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$2 across the reported FE$$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$3 cases, while VINO showed lower relative $$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$4 errors than residual PINO on anti-derivative, Poisson, and Darcy benchmarks and exhibited decreasing median error under mesh refinement (Eivazi et al., 27 Mar 2025, Eshaghi et al., 2024). In geophysical wave prediction, PICNO reduced relative $$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$5 error by 27.4%, 35.2%, and 53.1% at 8 Hz, 10 Hz, and 12 Hz, respectively, relative to a purely data-driven CNO, with inference time $$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$6 s versus $$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$7 s for a conventional solver (Ma et al., 22 Jul 2025).

Phase-field PINO exhibits similar trends. PF-PINO reported relative $$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$8 errors of $$\mathcal{P}[u](x,t;\mu)=0\quad \text{for }(x,t)\in\Omega\times[0,T],$$9 versus $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$0 for pencil-electrode corrosion, $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$1 versus $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$2 for electro-polishing corrosion, $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$3 versus $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$4 for dendritic solidification, and $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$5 versus $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$6 for spinodal decomposition, all comparing PF-PINO with conventional FNO (Chen et al., 10 Mar 2026). In electromagnetic inverse scattering, PINO with FNO, U-FNO, and F-FNO backbones consistently outperformed CSI-type methods in noisy, single-frequency, multi-frequency, and phaseless settings; for example, with-phase single-frequency FNO-PINO at 10% noise reported SSIM $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$7 and RMSE $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$8, versus CSI SSIM $$\mathcal{B}[u](x,t;\mu)=0,\qquad \mathcal{I}[u](x;\mu)=0.$$9 and RMSE $$\mathcal{G}_\theta:(\mu,f)\mapsto u_\theta(\cdot;\mu,f),$$0 (Dong et al., 26 Mar 2026).

The application range itself is informative. The framework has been used for aerospace-oriented parametric PDE families (Wang et al., 21 Jun 2025), wave propagation (Ma et al., 22 Jul 2025), wall-bounded turbulence (Zhao et al., 29 Apr 2026), power-system surrogates with over 30 times speedup compared to high-order ODE solvers (Karampinis et al., 7 Nov 2025), and multiscale FE$$\mathcal{G}_\theta:(\mu,f)\mapsto u_\theta(\cdot;\mu,f),$$1 constitutive surrogates (Eivazi et al., 27 Mar 2025). This suggests that PINO is functioning as a unifying computational pattern across domains where many-query solution evaluation is more important than solving a single PDE instance once.

6. Theory, misconceptions, and limitations

A central theoretical claim of the field is that the neural-operator backbone inherits approximation guarantees unavailable to standard PINNs. FNO is described as a universal approximator for continuous operators and discretization convergent in the limit of mesh refinement, which is why physics loss imposed on finer grids can improve high-frequency fidelity without sacrificing resolution transfer (Li et al., 2021). More generally, rigorous results establish approximation-error bounds for physics-informed operator learning and show that PINNs and physics-informed DeepONets or FNOs mitigate the curse of dimensionality for nonlinear parabolic PDEs under the assumptions stated in that framework (Ryck et al., 2022).

These guarantees do not remove the main practical limitations. Spectral differentiation assumes smoothness, so shocks and discontinuities remain difficult for residual-based PINO; the original PINO discussion explicitly notes that shocks may require artificial viscosity, shock-capturing, or weak-form residuals tailored to weak solutions (Li et al., 2021). Non-periodic domains and complex geometries are also nontrivial for Fourier layers, which is why wavelet, convolutional, spline, FE-based, or low-rank point-cloud operators recur throughout the literature (N et al., 2023, Ma et al., 22 Jul 2025, Schaffer et al., 9 Sep 2025). Aggressive spectral truncation can suppress high-frequency content, and LFR-PINO notes possible damping of sharp gradients and aliasing sensitivity when $$\mathcal{G}_\theta:(\mu,f)\mapsto u_\theta(\cdot;\mu,f),$$2 is chosen too small (Wang et al., 21 Jun 2025).

Another misconception is that physics-only training automatically yields the physically correct solution. Several studies report optimizer sensitivity, degenerate minima, or spurious solutions in purely physics-informed settings. Multi-stage curriculum training with optimizer resets was introduced precisely because single-stage unsupervised training could stagnate (Marcandelli et al., 2 Feb 2026), and hybrid supervision was required in the SPINONet Eikonal case to avoid degenerate convergence (Garg et al., 23 Mar 2026).

Finally, the framework is not limited to strong-form residual minimization. VINO shows that a physics-informed neural operator can be formulated by minimizing an energy functional on a discretized domain, and EquiNO shows that some constraints can be embedded hard at the architecture level through reduced bases rather than soft penalties (Eshaghi et al., 2024, Eivazi et al., 27 Mar 2025). A plausible implication is that future progress is likely to depend less on the phrase “physics-informed” itself than on how well the chosen operator basis, residual representation, and optimization schedule match the PDE class, boundary structure, and numerical regularity of the target problem.