Neural Operator Flows

• Neural Operator Flows are continuous-depth models that learn operators mapping functions, fields, or distributions between infinite-dimensional spaces.
• They leverage invertible mappings and flow-matching techniques to enable exact likelihood estimation and efficient Bayesian functional regression.
• These methods demonstrate strong performance in applications such as PDE control, lattice quantum field sampling, and stochastic process learning.

Neural operator flows are a family of continuous-depth, transport-based, or evolution-based constructions in which the learned object is an operator acting on functions, fields, or distributions rather than a finite-dimensional map on fixed-size vectors. In the most specific sense, “Neural Operator Flows” denotes the OpFlow framework for universal functional regression, where an invertible neural operator maps a non-Gaussian data function space to a Gaussian-process latent space and thereby supports exact likelihood estimation for functional point evaluations and posterior sampling (Shi et al., 2024). In a broader literature, the term also refers to operator-valued flows for stochastic process learning, Bayesian updating, lattice field theory sampling, and continuous-depth approximation on infinite-dimensional spaces (Shi et al., 7 Jan 2025, Chen et al., 2019, Máté et al., 2024, Chen et al., 21 May 2026).

1. Terminological scope and defining ideas

The phrase “neural operator flow” is not used uniformly across the literature. One line of work uses it to denote an invertible neural operator between function spaces, extending normalizing flows from finite-dimensional vectors to infinite-dimensional objects; this is the setting of OpFlow and universal functional regression (Shi et al., 2024). A second line uses neural operators as time-dependent vector fields in continuous normalizing flows or flow-matching systems on function spaces, with the aim of learning stochastic-process priors and tractable finite-dimensional marginals (Shi et al., 7 Jan 2025). A third line treats a neural operator as an evolution rule that transports distributions or fields through time, as in particle-based Bayesian updating and operator-based flows for lattice quantum field theory (Chen et al., 2019, Máté et al., 2024).

Across these variants, the shared structural idea is that the learned model acts on objects with function-space semantics. Inputs and outputs are typically fields, trajectories, point-evaluation collections, or probability measures induced by functions. This distinguishes neural operator flows from ordinary normalizing flows or neural ODEs that are tied to a single finite-dimensional ambient space.

A recurring motivation is that many scientific and statistical problems are intrinsically operator-valued. In universal functional regression, the objective is to learn a prior over non-Gaussian function spaces while retaining likelihood evaluation and posterior inference (Shi et al., 2024). In operator flow matching, the objective is to learn stochastic process priors on function spaces that provide the probability density of the values of any collection of points and enable functional regression at new points (Shi et al., 7 Jan 2025). In particle flow Bayes’ rule, Bayes updating is treated as an operator on distributions and realized as a continuous-time particle transport (Chen et al., 2019). In multi-lattice quantum field theory sampling, the lattice is viewed as a discretization of an underlying field theory, so the sampler is formulated as a discretization of a continuous operator flow between functional Boltzmann measures (Máté et al., 2024).

A common misconception is that neural operator flows are merely neural operators unrolled in time. The literature is more heterogeneous. Some methods are explicitly bijective and likelihood-based, some are simulation-free flow-matching systems, some are ODE-defined transports over particles or fields, and some are continuous-depth approximation frameworks whose discretizations recover ResNet-type or plain neural operator architectures (Shi et al., 2024, Shi et al., 7 Jan 2025, Chen et al., 21 May 2026).

2. Function-space formulations

A central formulation is the invertible operator pair used in OpFlow. The paper defines a forward operator

$\mathcal{G} : \mathcal{A} \to \mathcal{U}$

and an inverse operator

$$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$

where $\mathcal{U}$ is the data function space and $\mathcal{A}$ is a latent Gaussian-process space. In practice, the learned model is the inverse flow

$$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$

with $v^0=u$ and $v^t=a$ (Shi et al., 2024). This establishes a change-of-variables mechanism on function spaces rather than on vectors.

Operator flow matching instead starts from a real separable Hilbert space $\mathcal H$ and a time-dependent vector field

$G_t : \mathcal H \to \mathcal H,$

driving the ODE

$$\frac{\partial \Phi_t(h_0)}{\partial t} = G_t(\Phi_t(h_0)), \qquad \Phi_0(h_0)=h_0.$$

The corresponding probability path satisfies the continuity equation

$$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$0

so the learned operator field transports a reference measure $$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$1 toward a target measure $$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$2 in function space (Shi et al., 7 Jan 2025).

Particle Flow Bayes’ Rule adopts a related but distribution-conditional view. Given particles

$$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$3

and a new observation $$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$4, each particle evolves according to

$$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$5

The induced density $$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$6 satisfies

$$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$7

and along trajectories

$$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$8

The paper’s existence argument connects this deterministic transport to Langevin dynamics and shows that a unified Bayes-flow operator can be obtained through open-loop/closed-loop equivalence (Chen et al., 2019).

In multi-lattice quantum field theory, the target objects are functional Boltzmann measures rather than stochastic-process priors or particle posteriors. The model introduces a time-dependent operator $$\mathcal{F} = \mathcal{G}^{-1} : \mathcal{U} \to \mathcal{A},$$9 whose integral transports the free-theory measure to the interacting $\mathcal{U}$0 measure. After discretization on a chosen lattice, $\mathcal{U}$1 induces an ordinary time-dependent vector field $\mathcal{U}$2, so the discretized operator becomes a continuous normalizing flow over lattice field configurations (Máté et al., 2024).

This suggests a useful unifying description: neural operator flows typically specify either an invertible map between function spaces or a time-dependent operator field whose induced trajectory transports functions, fields, or distributions through a continuous path. The exact object being transported differs by application.

3. Exact likelihoods, stochastic-process priors, and functional regression

The most explicit probabilistic use of neural operator flows appears in universal functional regression. OpFlow aims to learn a prior over non-Gaussian function spaces that remains mathematically tractable for regression. Its exact log-likelihood for a sample $\mathcal{U}$3 observed on a discretization $\mathcal{U}$4 is written as

$\mathcal{U}$5

where $\mathcal{U}$6 lies in the Gaussian-process latent space (Shi et al., 2024). Because the latent prior is Gaussian, posterior inference can be performed in that space and mapped back by $\mathcal{U}$7.

The training objective in OpFlow combines negative log-likelihood with a $\mathcal{U}$8-Wasserstein regularizer during warmup,

$\mathcal{U}$9

followed by a finetuning stage with $\mathcal{A}$0 (Shi et al., 2024). Posterior sampling is performed with stochastic gradient Langevin dynamics in latent GP space: $\mathcal{A}$1

Operator flow matching targets a similar endpoint—tractable stochastic-process learning—but replaces invertibility-based training with flow matching. Its conditional path uses Gaussian interpolants between $\mathcal{A}$2 and $\mathcal{A}$3,

$\mathcal{A}$4

with conditional vector field

$\mathcal{A}$5

The model is trained using conditional flow matching,

$\mathcal{A}$6

and provides the density of any finite collection of points through the log-density evolution

$\mathcal{A}$7

(Shi et al., 7 Jan 2025).

For functional regression, OFM writes the posterior over values at $\mathcal{A}$8 points conditioned on noisy observations at $\mathcal{A}$9 points as

$$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$0

and uses SGLD for posterior sampling (Shi et al., 7 Jan 2025).

Empirically, OpFlow is reported to match exact GP regression well on Gaussian-process data, to learn truncation constraints on truncated Gaussian processes, and to outperform standard GP on seismic waveforms from KiK-net (Shi et al., 2024). OFM reports strong results on 1D and 2D Gaussian processes, truncated Gaussian processes, and complex fields such as 2D Navier–Stokes, black hole simulations, and MNIST signed-distance functions; for example, Table 1 reports OFM SMSE $$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$1 and MSLL $$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$2 on 1D GP, SMSE $$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$3 and MSLL $$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$4 on 2D GP, and mean error $$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$5 with std error $$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$6 on 1D truncated GP (Shi et al., 7 Jan 2025).

4. Operator-flow models for dynamical systems, PDEs, and control

In PDE-oriented work, “neural operator flow” often refers to a learned evolution operator that advances fields or latent states. The emphasis is typically not exact likelihood, but surrogate modeling, discretization transfer, or control.

A representative example is the use of Neural Implicit Flow as a mesh-agnostic neural operator for latent dynamics of the Kuramoto–Sivashinsky, forced Korteweg–de Vries, and Sine-Gordon equations. The model combines ParameterNet and ShapeNet, with ParameterNet generating the parameters of the decoder-like ShapeNet. The study concludes that NIF is often more accurate as a surrogate, while DeepONet’s latent space is more interpretable on the examples considered (Nasim et al., 2024). Reported reconstruction errors include $$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$7 for DeepONet versus $$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$8 for NIF on KS bursting dynamics, $$\mathcal{F}_\theta := \mathcal{F}^t_\theta \circ \mathcal{F}^{t-1}_\theta \circ \cdots \circ \mathcal{F}^0_\theta,$$9 versus $v^0=u$0 on fKdV traveling wave data, and $v^0=u$1 versus $v^0=u$2 on SG (Nasim et al., 2024).

For fluid control, physics-informed neural-operator predictive control formulates wall-bounded turbulence control as model-based reinforcement learning with jointly trained observer and policy operators. The observer predicts the interior flow field from wall actuation, the policy predicts wall-normal blowing and suction from boundary pressure, and both are implemented with neural operators, especially FNOs (Zhao et al., 3 Oct 2025). The framework reports 43.5% drag reduction on flows with Reynolds numbers not included in training and 39.0% drag reduction at $v^0=u$3, with the abstract stating that this outperforms previous fluid control methods by more than 32% in the high-Reynolds-number setting (Zhao et al., 3 Oct 2025).

In large-eddy simulation, HUFNO treats coarse-flow evolution as an operator from previous coarse states to the next state increment. The model uses Fourier neural operator blocks in the periodic directions and a U-Net in the non-periodic wall-normal direction. It is evaluated on 3D turbulent flow over periodic hills, where the original FNO diverges at $v^0=u$4 and HUFNO yields better predictions of mean velocity, Reynolds stresses, energy spectra, wall shear, and separation structures than FNO, U-Net, SMAG, and WALE (Wang et al., 17 Apr 2025). The reported times for 10,000 DNS time steps are $v^0=u$5 s, $v^0=u$6 s, and $v^0=u$7 s for HUFNO at $v^0=u$8, versus $v^0=u$9 s, $v^t=a$0 s, and $v^t=a$1 s for SMAG and $v^t=a$2 s, $v^t=a$3 s, and $v^t=a$4 s for WALE (Wang et al., 17 Apr 2025).

Other PDE surrogates emphasize operator learning over geometries or discontinuities rather than explicit continuous-depth transport. Fusion-DeepONet addresses geometry-dependent hypersonic and supersonic flows by fusing branch-network hidden-layer outputs into multiple trunk layers; on the semi-ellipse blunt-body problem it reports overall $v^t=a$5 errors of 3.70 on the uniform grid and 6.39 on irregular grids, compared with 4.54 for U-Net and 8.21 for FNO on the uniform grid and 49.46 for MeshGraphNet on irregular grids (Peyvan et al., 3 Jan 2025). RiemannONets study neural operators for 1D compressible Euler Riemann problems with pressure jumps up to $v^t=a$6; the paper reports that U-Net performs best in the extreme LeBlanc regime, while a two-step SVD-based DeepONet improves robustness and interpretability through a hierarchical orthonormal basis (Peyvan et al., 2024).

These developments indicate that the “flow” in neural operator flow can denote at least three distinct notions in scientific computing: a learned continuous-time transport, a neural-operator-defined latent dynamics, or a learned field evolution operator used autoregressively or in closed-loop control.

5. Architectural patterns and training strategies

Several architectural motifs recur across the literature.

A first motif is function-space coupling with tractable Jacobians. In OpFlow, each invertible layer contains actnorm, a partitioning strategy, and an affine coupling layer driven by an FNO block: $v^t=a$7

$v^t=a$8

The choice between domain partitioning and codomain partitioning leads to an expressivity-artifact tradeoff: domain partitioning is more expressive but can introduce checkerboard artifacts in super-resolution, whereas codomain partitioning is artifact-free in super-resolution but somewhat less expressive (Shi et al., 2024).

A second motif is operator-valued vector fields with resolution-agnostic backbones. OFM uses a neural operator, specifically an FNO, to parameterize the vector field $v^t=a$9 on collections of point evaluations. The divergence term needed for likelihood estimation is approximated with the Hutchinson trace estimator,

$\mathcal H$0

reducing the complexity from quadratic to linear in the number of points (Shi et al., 7 Jan 2025).

A third motif is set-conditioned or observation-conditioned flows. PFBR summarizes the current posterior particle cloud by a DeepSets-like embedding,

$\mathcal H$1

and conditions the vector field on this representation, the new observation, the current particle, and time: $\mathcal H$2 (Chen et al., 2019).

A fourth motif is neural operators designed for discretization transfer. In multi-lattice quantum field theory, the operator is defined continuously and then discretized on any chosen lattice. The base distribution is not a standard Gaussian but the free theory $\mathcal H$3, whose covariance is diagonal in momentum space, making it physically aligned with the target theory (Máté et al., 2024). The authors report that training on a curriculum of smaller lattices before the $\mathcal H$4 target lattice yields comparable final performance with about 2.4 times faster training than training directly on the target lattice (Máté et al., 2024).

A fifth motif is two-stage or warmup-plus-finetune training. OpFlow first optimizes likelihood plus $\mathcal H$5 regularization and then finetunes with likelihood alone (Shi et al., 2024). OFM performs prior learning by coupling reference and data minibatches with minibatch optimal transport and regressing to straight-line conditional displacements (Shi et al., 7 Jan 2025). PFBR is trained meta-learnedly across many Bayesian inference tasks with a KL/cross-entropy-style objective over sequential posterior updates (Chen et al., 2019).

6. Theory, universality, and limitations

The theoretical literature has increasingly formalized neural operator flows as continuous-depth approximation mechanisms on infinite-dimensional spaces. An abstract framework introduced in 2026 defines neural flows on a separable Hilbert latent space $\mathcal H$6 through

$\mathcal H$7

and approximates an operator $\mathcal H$8 by

$\mathcal H$9

The paper proves well-posedness and universal approximation results for both composition and separation flow structures and states, “to the best of our knowledge,” the first universal approximation result for flow-based models between infinite-dimensional spaces (Chen et al., 21 May 2026).

The composition structure,

$G_t : \mathcal H \to \mathcal H,$0

recovers a ResNet-type discretization,

$G_t : \mathcal H \to \mathcal H,$1

while the separation structure,

$G_t : \mathcal H \to \mathcal H,$2

yields a plain architecture under splitting-based discretization (Chen et al., 21 May 2026). The same paper also provides universal approximation results for convolutional neural flow models.

A related mathematical synthesis studies neural operators as mappings $G_t : \mathcal H \to \mathcal H,$3 and interprets iterative neural-operator updates as approximate gradient flows in function space,

$G_t : \mathcal H \to \mathcal H,$4

Under this interpretation, the paper proves stability bounds, contraction-based convergence, and clustering toward critical points in Sobolev norm (Le et al., 2024). This suggests a continuous-time reading of operator iteration even when a model is not explicitly built as a neural ODE.

The literature also states clear limitations. OpFlow notes that codomain partitioning can be harder to train and less expressive, whereas domain decomposition can cause checkerboard or jagged artifacts in zero-shot super-resolution; it also remarks that, although the theory supports irregular grids, the FNO-based implementation operates on regular grids (Shi et al., 2024). OFM relies on assumptions including a Gaussian reference measure with trace-class covariance and target support on the associated Cameron–Martin space (Shi et al., 7 Jan 2025). The multi-lattice sampling paper shows that operator-based flows generalize to nearby unseen lattice sizes but do not extrapolate strongly to much larger sizes than those seen during training (Máté et al., 2024). PFBR notes that training can be expensive and introduces mini-batch embedding and sequence segmentation to reduce cost (Chen et al., 2019).

A broader interpretive caution follows from this body of work. “Neural operator flows” is best understood as an umbrella concept rather than a single architecture. In one setting it denotes an invertible functional prior, in another a function-space flow-matching model, in another a transport-based Bayesian update operator, and in another a continuous-depth approximation principle. What unifies these formulations is the use of operator learning to define transports, trajectories, or evolutions at the level of functions, fields, or measures rather than fixed finite-dimensional arrays.