Physics-informed Neural Operators

• Physics-informed Neural Operators (PINOs) are operator-learning models that incorporate PDE residuals and boundary conditions to learn mappings across function spaces.
• They combine coarse data with fine-grid physics constraints, outperforming data-only approaches on canonical PDE benchmarks.
• PINO methodologies leverage diverse backbones like FNO, DeepONet, and CViT to enhance accuracy, generalization, and efficient many-query inference.

Searching arXiv for recent and foundational PINO papers to ground the article. Physics-informed Neural Operators (PINOs) are operator-learning models trained with physics-informed supervision to approximate solution operators of governing differential equations. Rather than learning a single solution instance, a PINO learns a mapping between function spaces—for example, from coefficients, geometry, initial or boundary data, or sparse measurements to full solution fields—while enforcing PDE residuals, boundary or initial conditions, conservation constraints, or variational structure during training. In this sense, PINOs occupy a middle position between data-driven neural operators and Physics-Informed Neural Networks (PINNs): they retain cross-instance generalization and amortized inference, but inject physical structure directly into optimization (Li et al., 2021, Dai et al., 20 Jan 2026).

1. Conceptual foundations

The defining viewpoint is operator learning. Given an input space $X$ and an output space $Y$, a neural operator learns a map $G: X \to Y$ such that $y = G(u)$ approximates the solution of a PDE or reduced-order model for new inputs. In coronary flow modeling, this is stated explicitly as a map from geometric descriptors, boundary functions, and material parameters to pressure and flow fields; in the general PDE setting, the same idea is used for coefficients, forcings, initial conditions, and boundary data (Zhu et al., 17 Feb 2026).

This distinguishes PINOs from classical PINNs. A PINN typically parameterizes one solution $u_\theta(x,t)$ for one fixed PDE instance and minimizes residuals at collocation points. A PINO instead parameterizes the solution operator itself across a family of instances. The unifying survey on PINNs and neural operators formulates this difference along three axes: what is learned, how physics enters, and how computation is amortized. PINNs learn single-instance solutions and optimize per instance; neural operators learn family-level operators and amortize computation across instances; PINOs combine the latter with physics-based supervision (Dai et al., 20 Jan 2026).

The 2021 foundational formulation of PINO emphasized a further distinction from purely supervised neural operators: coarse-resolution training data can be combined with higher-resolution PDE residuals. In that formulation, PINO was presented as a hybrid approach that uses coarse data for optimization guidance while enforcing PDE constraints on finer grids, which in turn supports zero-shot super-resolution and physics-only operator learning (Li et al., 2021).

2. Operator parameterizations and physics-informed objectives

Early PINOs were built on the Fourier Neural Operator (FNO), whose spectral layer can be written as

$$\phi_{\ell+1}(x)=\sigma\!\left(W_\ell \phi_\ell(x)+\mathcal{F}^{-1}\!\big(R_\ell(k)\cdot \mathcal{F}[\phi_\ell](k)\big)\right),$$

with learned spectral multipliers $R_\ell(k)$ and pointwise linear maps $W_\ell$. This gives a global, FFT-based operator parameterization with discretization-convergent behavior in the mesh-refinement limit, and it underlies the first PINO construction (Li et al., 2021, Zhu et al., 17 Feb 2026).

The physics-informed part enters through composite losses. A representative PINO objective augments data mismatch with residual, boundary, and initial penalties,

$$\mathcal{L} = \mathcal{L}_{\mathrm{data}} + \lambda \|\mathcal{R}_{\mathrm{PDE}}\|^2 + \mu \|\mathcal{B}_{\mathrm{BC}}\|^2 + \alpha \|\mathcal{I}_{\mathrm{IC}}\|^2 + \beta\,\mathcal{R}_{\mathrm{reg}},$$

where the residual may encode full-order or reduced-order governing equations, and the regularization term may encode smoothness, monotonicity, or other priors (Zhu et al., 17 Feb 2026). In the original PINO work, this same principle was instantiated as coarse data loss plus fine-grid PDE, boundary, and initial-condition losses, allowing operator training with partial or even zero paired data (Li et al., 2021).

Although FNO remains the best-known backbone, later work makes clear that PINO is not architecture-specific. Comparative training studies evaluate DeepONet, FNO, and Continuous Vision Transformer (CViT) backbones under the same physics-informed regime, while application papers use graph operators, Transformer-KAN Neural Operators, Transolver-based operators, and hypernetwork-based constructions such as HyPINO (Chen et al., 4 Jun 2026, Ding et al., 24 Nov 2025, Wang et al., 6 May 2026, Bischof et al., 5 Sep 2025). The common feature is not the spectral layer itself, but the use of operator learning with physics-based supervision.

Weak-form and variational PINOs are a notable branch. The coronary FFR review highlights weak-form residuals as a stabilization device, especially in Bi-variational Physics-Informed Operator Networks, and the general training study likewise reports that training robustness depends strongly on residual formulation, derivative evaluation, and collocation design (Zhu et al., 17 Feb 2026, Chen et al., 4 Jun 2026).

3. Training pathologies and algorithmic remedies

PINO training inherits many of the optimization difficulties known from PINNs, but in an operator-learning setting. A 2026 empirical study shows that gradient conflicts and causal violation arise naturally in PINOs as well, and that mitigation strategies from PINN training remain effective. The same study reports consistently strong performance for CViT backbones, and identifies optimizer choice, loss balancing, and collocation-point sampling as central determinants of robustness (Chen et al., 4 Jun 2026).

Several remedies have been proposed for distinct failure modes. For nonperiodic problems, the Fourier representation itself becomes a bottleneck because padded Fourier extensions induce ill-conditioning, Gibbs artifacts, and unstable high-order derivatives. “Fourier Continuation for Exact Derivative Computation in Physics-Informed Neural Operators” addresses this by constructing smooth periodic continuations before spectral differentiation. In the reported 1D blowup experiments, the best FC-PINO strategy improves equation loss by several orders of magnitude relative to padded PINO and accurately captures third-order derivatives of nonsmooth solutions (Maust et al., 2022).

Symmetry-informed loss design is another line of work. “Generalized Lie Symmetries in Physics-Informed Neural Operators” argues that naïve point-symmetry augmentation can yield no useful training signal, because the added term often vanishes identically or merely rescales the base PDE residual. The proposed alternative uses evolutionary representatives of generalized symmetries, producing loss terms that behave like Sobolev-type penalties on the residual and improve data efficiency on Burgers and Darcy benchmarks (Wang et al., 1 Feb 2025).

Two further directions broaden training beyond the static in-distribution regime. “Can Physics Informed Neural Operators Self Improve?” shows that self-training with pseudo-labels narrows the gap between physics-only PINO and data-plus-physics training, reaching $1.07\times$ the Burgers error and $Y$0 the Darcy error of FNOs trained with both data and physics (Majumdar et al., 2023). “Replay-Based Continual Learning for Physics-Informed Neural Operators” extends PINOs to out-of-distribution adaptation by combining residual-based replay selection, distillation, and LoRA within a Transolver-based operator, mitigating catastrophic forgetting while remaining fully physics-informed and label-free (Wang et al., 6 May 2026).

4. Application domains and problem classes

The published literature now spans both canonical PDE benchmarks and highly domain-specific scientific workflows. The table below summarizes representative patterns already reported.

Domain	PINO formulation	Representative finding
Parametric PDE benchmarks	FNO-based PINO on Burgers, Darcy, Navier–Stokes/Kolmogorov	Physics-only operator learning and zero-shot super-resolution (Li et al., 2021)
Phase-field and microstructure	Coupled Allen–Cahn/Cahn–Hilliard PINO; PF-PINO for corrosion, solidification, spinodal decomposition	Improved long-term stability and generalization (Gangmei et al., 24 Jul 2025, Chen et al., 10 Mar 2026)
MHD and plasma equilibrium	Tensor-FNO PINO for incompressible MHD; PINO for nonlinear Grad–Shafranov	Accurate laminar MHD for $Y$1; robust semi-supervised equilibrium inference (Rosofsky et al., 2023, Ding et al., 24 Nov 2025)
Inverse and measurement-driven problems	Ocean-wave reconstruction, inverse scattering, thermoelectric property inference, speech analysis, coronary FFR	Label-free or sparse-label inference under physics constraints (Ehlers et al., 5 Aug 2025, Dong et al., 26 Mar 2026, Moon et al., 9 Jun 2025, Yokota et al., 21 Jun 2026, Zhu et al., 17 Feb 2026)

In canonical operator-learning benchmarks, PINO was first demonstrated on Burgers, Darcy, and Navier–Stokes/Kolmogorov flow. There it reduced generalization error relative to data-only FNO, enabled physics-only training when labels were unavailable, and supported instance-wise fine-tuning with anchor losses for difficult dynamical regimes (Li et al., 2021).

Phase-field modeling has become a major testbed because of stiff, high-order, and coupled dynamics. A PINO for coupled Allen–Cahn and Cahn–Hilliard equations in Al–Cu precipitation uses pseudo-spectral differentiation and reports that replacing finite differences with Fourier derivatives improves the Cahn–Hilliard equation loss by twelve orders of magnitude. PF-PINO, a later phase-field framework, extends the idea to electrochemical corrosion, dendritic solidification, and spinodal decomposition, and reports substantially lower relative errors and better long-horizon behavior than conventional FNO (Gangmei et al., 24 Jul 2025, Chen et al., 10 Mar 2026).

The same operator-learning logic has been transferred to plasma and MHD settings. A tensor-FNO PINO reproduces two-dimensional incompressible magnetohydrodynamics accurately for laminar flows with $Y$2, but loses fidelity as magnetic energy shifts toward higher wavenumbers in more turbulent regimes (Rosofsky et al., 2023). For the nonlinear Grad–Shafranov equation, a semi-supervised PINO trained with sparse labeled interior points and physics constraints reaches $Y$3 interpolation error and $Y$4 extrapolation error on out-of-distribution samples, with millisecond-scale inference after TensorRT optimization (Ding et al., 24 Nov 2025).

Measurement-driven and inverse settings are especially revealing because they expose whether operator learning remains useful when the inputs are sparse, indirect, or noisy. A PINO for ocean waves reconstructs phase-resolved nonlinear wave fields from buoy time series or radar snapshots without ground-truth wavefields during training (Ehlers et al., 5 Aug 2025). In electromagnetic inverse scattering, PINO jointly optimizes neural-operator parameters and a dielectric tensor under state, data, and total-variation losses, covering with-phase, phaseless, single-frequency, and multi-frequency settings within one differentiable framework (Dong et al., 26 Mar 2026). A thermoelectric PINO uses sparse field measurements to infer temperature-dependent thermal conductivity and Seebeck coefficient across unseen materials in a label-free manner (Moon et al., 9 Jun 2025). A speech-production PI-DeepONet predicts the periodic coupled solution of a vocal-fold–vocal-tract system from tract shapes without supervised field labels (Yokota et al., 21 Jun 2026). In imaging-derived coronary FFR, graph-based and bi-variational operator-like models are used to output pressure, flow, and whole-vessel pullback curves under conservation and boundary-consistency constraints (Zhu et al., 17 Feb 2026).

5. Empirical behavior, accuracy, and deployment

Reported performance is strongly task-dependent, but the literature provides several stable reference points. In the foundational PINO study, Burgers operator learning trained on $Y$5 data with physics at $Y$6 achieves relative errors of $Y$7, $Y$8, and $Y$9 at $G: X \to Y$0, $G: X \to Y$1, and $G: X \to Y$2 resolutions, respectively, while Darcy operator learning trained at $G: X \to Y$3 with physics at $G: X \to Y$4 yields $G: X \to Y$5, $G: X \to Y$6, and $G: X \to Y$7 at the corresponding resolutions (Li et al., 2021). These are among the clearest demonstrations that physics-informed operator learning can improve high-resolution behavior relative to data-only neural operators.

Several later application papers translate this into deployment-oriented metrics. In angiography-derived coronary FFR, a Bi-variational Physics-Informed Operator Network reports ischemia-detection performance of $G: X \to Y$8, $G: X \to Y$9, and $y = G(u)$0, with inference time below $y = G(u)$1 versus approximately $y = G(u)$2 for steady or pseudotransient CFD and approximately $y = G(u)$3 for a PINN baseline (Zhu et al., 17 Feb 2026). In fusion equilibrium reconstruction, semi-supervised PINO shows a degradation factor of $y = G(u)$4 under extrapolation, compared with $y = G(u)$5 for supervised models, while retaining $y = G(u)$6 latency with TensorRT (Ding et al., 24 Nov 2025). In speech production analysis, the reported errors are $y = G(u)$7 for glottal volume flow and $y = G(u)$8 for speech waveforms (Yokota et al., 21 Jun 2026). In PF-PINO, electro-polishing corrosion improves from $y = G(u)$9 relative $u_\theta(x,t)$0 error under FNO to $u_\theta(x,t)$1 under PF-PINO, while dendritic solidification improves from $u_\theta(x,t)$2 to $u_\theta(x,t)$3 (Chen et al., 10 Mar 2026).

These results are not directly comparable across domains, but they support a common pattern. PINOs tend to be most attractive when three conditions hold simultaneously: the forward physics is expensive, labels are sparse or absent, and the target workflow requires repeated evaluation across varying parameters, geometries, or boundary conditions. This suggests that the main empirical advantage is not merely accuracy at one operating point, but the combination of physics-aware regularization with amortized many-query inference.

6. Limitations, misconceptions, and research directions

A common misconception is that PINOs are simply PINNs equipped with FFT layers. The literature instead treats them as operator learners: what changes is not only the architecture, but the object being learned and the computational load being amortized across problem instances (Dai et al., 20 Jan 2026). A second misconception is that physics-informed training alone guarantees out-of-distribution correctness. Multiple studies report the opposite: nonperiodicity can destabilize spectral differentiation, gradient conflicts and causal violation can derail training, FNO backbones can fail under non-periodic wall-bounded conditions, and high-Re or high-frequency dynamics can outrun the operator’s retained spectral capacity (Maust et al., 2022, Chen et al., 4 Jun 2026, Rosofsky et al., 2023).

Boundary and geometry handling remain central bottlenecks. FNO-based PINOs are naturally aligned with periodic domains, but several applications—lid-driven cavity flow, coronary trees, inverse scattering, speech waveguides, free-surface problems, and complex phase-field interfaces—require nonperiodic treatments, graph encodings, coordinate decoders, or special continuation strategies. This is why later work moves toward graph operators, DeepONet-like coordinate decoders, CViT, Transolver, and hypernetwork constructions rather than relying exclusively on spectral convolutions (Chen et al., 4 Jun 2026, Wang et al., 6 May 2026, Bischof et al., 5 Sep 2025).

Deployment raises a separate set of concerns. The clinical FFR review explicitly emphasizes calibration, uncertainty quantification, residual-based confidence, and quality-control gatekeeping as necessary for safe use, and the broader survey of PINNs and neural operators similarly argues for proposal–correct–validate workflows rather than unconditional trust in the learned surrogate (Zhu et al., 17 Feb 2026, Dai et al., 20 Jan 2026). This suggests that PINOs are best viewed as physics-regularized surrogate operators whose outputs should still be checked against residuals, constraints, and domain-specific quality controls.

Current research directions are correspondingly diverse. They include higher-dimensional Fourier continuation for nonperiodic problems, continual learning under domain shift, mixed supervision with sparse labels plus physics, residual-driven iterative refinement, better handling of coupled multiphysics and irregular geometries, and broader benchmark standardization for calibration, uncertainty, and long-horizon stability (Maust et al., 2022, Wang et al., 6 May 2026, Bischof et al., 5 Sep 2025, Chen et al., 4 Jun 2026). Across these lines of work, the central theme remains stable: PINOs are not a single architecture, but a training paradigm for learning PDE solution operators under explicit physical structure.