GeoFunFlow: Geometry-Conditioned Field Inference

• GeoFunFlow is a geometric diffusion model framework that maps complex geometries and sparse sensor data to a posterior distribution over physical fields.
• It integrates a latent autoencoder and diffusion processes to reconstruct continuous field representations from irregular discretizations.
• The approach improves reconstruction accuracy and uncertainty quantification, outperforming traditional CNN and VAE baselines, even with high noise levels.

Searching arXiv for GeoFunFlow and closely related papers to ground the article in the relevant literature. GeoFunFlow denotes a family of geometry-conditioned generative methods for field inference on complex domains. In its most explicit formulation, GeoFunFlow is a geometric diffusion model framework for inverse problems on complex geometries that maps a conditioning instance consisting of geometry, sensor locations, and sparse noisy observations to a posterior distribution over physical fields on that geometry (Wang et al., 28 Sep 2025). In adjacent usage, the term also describes the broader idea of a geometry-to-flow map in computational fluid dynamics, where obstacle geometry is used to predict entire flow fields as functional outputs (Hu et al., 2024). A subsequent extension, GeoFunFlow-3D, adapts the framework to forward, zero-observation 3D aerodynamic inference and adds explicit physics guidance, flow matching, and topology-aware refinement for high-fidelity 3D fields over complex geometries (Jiang et al., 25 Apr 2026).

1. Conceptual scope and problem formulation

GeoFunFlow is centered on learning mappings from geometry-conditioned inputs to spatially distributed physical fields. In the inverse-operator formulation, the target domain is not a single fixed grid but a family of geometries

$$\Lambda = \{\Omega\}, \quad \Omega \subset \mathbb{R}^d,$$

with varying shapes and irregular discretizations such as unstructured meshes and point clouds. On each geometry $\Omega$, the unknown state is a field

$$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$

and observations are sparse and noisy point samples

$$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$

The conditioning instance is

$$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$

and the learning objective is a conditional probabilistic operator

$$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$

that is, a posterior over fields conditioned on geometry and observations (Wang et al., 28 Sep 2025).

A closely related use of the term appears in geometry-to-flow prediction for 2D incompressible flow past obstacles. There, the problem is posed as

$$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$

where $u,v$ are velocity components, $p$ is pressure, and $\omega$ is vorticity. The outputs are instantaneous snapshots rather than time-averaged fields (Hu et al., 2024). This suggests that GeoFunFlow can be interpreted at two levels: as a specific inverse-operator framework on irregular geometries, and as a broader geometry-conditioned field generation paradigm.

The 3D extension shifts from inverse inference to pure forward generation. GeoFunFlow-3D is described as a forward, zero-observation extension in which inference is conditioned only on geometry, and on operating conditions if present, with targets including pressure, density, temperature, and velocity fields over complex 3D geometries (Jiang et al., 25 Apr 2026).

2. Antecedents in geometry-conditioned generative flow prediction

A concrete precursor to GeoFunFlow is the geometry-to-flow diffusion model for 2D incompressible flow around isolated obstacles (Hu et al., 2024). That model takes obstacle shape as input and predicts a full instantaneous field around the obstacle. The paper frames the task as learning a conditional distribution

$\Omega$0

where $\Omega$1 encodes obstacle geometry as a binary mask on a $\Omega$2 grid, and $\Omega$3 contains $\Omega$4 on the same grid.

The forward diffusion process uses a Markov chain

$\Omega$5

with $\Omega$6 and

$\Omega$7

The reverse process is conditioned on geometry,

$\Omega$8

and is implemented with a U-Net denoiser that predicts noise under the standard DDPM objective (Hu et al., 2024).

The architecture is a 2D U-Net with five encoder stages, mirrored decoding, skip connections, and cross-attention in the last two decoder stages. Geometry is injected as a prompt: the binary geometry field is encoded by two small neural networks into latent prompt vectors whose channel sizes match the last and second-to-last decoder layers. Cross-attention then combines decoder features $\Omega$9 with geometry prompt $$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$0 via

$$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$1

followed by

$$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$2

According to the description, this conditioning empirically “balances interpolation and extrapolation,” especially for complex shapes (Hu et al., 2024).

The training data are 2D incompressible Navier–Stokes simulations in OpenFOAM with free-stream velocity $$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$3, viscosity $$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$4, density $$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$5, and Reynolds number $$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$6, on body-fitted meshes of about 55,000 cells. The training geometries are circles, ellipses, rectangles, and triangles; test geometries include parallelograms, a cross, and the characters “PKU” (Hu et al., 2024).

The reported comparison is against a CNN/U-Net baseline with the same backbone but direct regression, and the abstract also mentions a VAE baseline. The diffusion model is described as superior for predicting instantaneous flow fields and handling complex geometries, with lower normalized $$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$7 velocity error and lower divergence metric than the CNN, especially for low-roundness shapes (Hu et al., 2024). A central interpretation is that deterministic regression tends to collapse toward a mean or time-averaged-like field, whereas conditional generative modeling can represent multiple plausible instantaneous realizations.

3. GeoFunFlow for inverse operator learning on complex geometries

The 2025 GeoFunFlow formulation generalizes geometry-conditioned field generation from fixed Cartesian CFD settings to inverse problems on families of irregular geometries (Wang et al., 28 Sep 2025). Its purpose is Bayesian field reconstruction under sparse and noisy observations, where classical PDE-constrained optimization would require repeated forward solves, adjoints, and posterior sampling loops on complex meshes.

The framework has two stages.

First, GeoFAE (Geometric Function Autoencoder) encodes a discretized field on an irregular mesh into a compact latent vector and decodes that latent representation into a continuous field $$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$8 evaluable at arbitrary coordinates. For a geometry $$u \in \mathcal{U}(\Omega) := H^s(\Omega; \mathbb{R}^p), \quad s \ge 1,$$9 discretized as a point cloud $$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$0, and a sensor subset $$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$1, the conditioning features at node $$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$2 are

$$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$3

where $$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$4. The autoencoder is

$$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$5

Second, a latent diffusion model trained via rectified flow learns a posterior over latent variables conditioned on sparse and noisy observations. If $$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$6 is the GeoFAE embedding of the full reference field and $$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$7 is the embedding of the partial observation, then with $$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$8 and $$X = \{x_j\}_{j=1}^m \subset \Omega,\qquad \mathbf{y} = \mathcal{S}_{\Omega,X}[u] + \epsilon,\quad \epsilon\sim\mathcal{N}(0,\Sigma).$$9,

$$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$0

and a Diffusion Transformer $$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$1 is trained with

$$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$2

At inference, the ODE

$$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$3

is integrated to obtain $$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$4, which is decoded into a continuous field sample (Wang et al., 28 Sep 2025).

The paper emphasizes three intended properties: reconstruction accuracy, uncertainty quantification, and efficient posterior sampling. Because the flow is defined in latent space rather than over variable-size raw discretizations, sampling requires integrating a low-dimensional ODE and avoids direct diffusion over large point clouds or meshes. The reported ablations further state that even one integration step is competitive, while about ten steps are effectively converged relative to a thousand-step reference (Wang et al., 28 Sep 2025).

4. Architectural components and geometric representation

GeoFAE is explicitly geometry-aware but mesh-agnostic. Its encoder uses a Perceiver to compress a variable-size set of mesh nodes into a fixed-size latent representation. Coordinates $$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$5 are first embedded by random Fourier features,

$$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$6

where the rows of $$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$7 are sampled from $$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$8. These are concatenated with embeddings of the mask and observed value features to form node embeddings $$\mathbf{c} = (\Omega, X, \mathbf{y}) \in \mathcal{C},$$9. Trainable latent queries $$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$0 then attend to the variable-sized node set through Perceiver cross-attention, followed by latent Transformer blocks, and are finally pooled into a vector $$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$1 (Wang et al., 28 Sep 2025).

The decoder is a CViT-style coordinate-based decoder. Query coordinates are Fourier-embedded and refined through repeated cross-attention with the latent tokens $$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$2, after which an MLP maps each refined query embedding to the target field dimension. This architecture makes the reconstruction continuous in coordinates and decouples the representation from any particular mesh resolution (Wang et al., 28 Sep 2025).

GeoFunFlow-3D adopts a different geometry-processing pipeline tailored to complex 3D aerodynamic settings. Each point $$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$3 is assigned a 9D feature

$$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$4

combining coordinates, surface normal, signed distance, principal curvature, and a reserved channel. A Graph Neural Operator is applied on a k-NN graph,

$$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$5

and the resulting unordered point features are projected onto a regular latent grid by Nadaraya-Watson kernel regression. This produces a geometry-aware latent representation used as the target manifold for flow matching (Jiang et al., 25 Apr 2026).

A distinct but terminologically adjacent antecedent is GFNet, whose full name is Geometric Flow Network for multi-view LiDAR semantic segmentation (Qiu et al., 2022). That work does not use the term GeoFunFlow explicitly, but it introduces a Geometric Flow Module that aligns and propagates features between range-view and bird’s-eye-view representations using exact geometric correspondences via the underlying 3D points. The paper interprets this as an align-before-fuse strategy, not late fusion. A plausible implication is that the “GeoFunFlow” label, when used more broadly, inherits part of its semantics from this earlier notion of geometry-constrained feature flow (Qiu et al., 2022).

5. Physics, sampling, and physical consistency

The 2025 GeoFunFlow inverse framework is explicitly described as data-driven rather than physics-informed. No PDE residuals or boundary-condition penalties are enforced during training or sampling. Physical structure is learned implicitly through training data generated by forward PDE solvers, and the method returns posterior samples by decoding latent ODE trajectories (Wang et al., 28 Sep 2025).

By contrast, the 2D geometry-to-flow diffusion precursor is also not explicitly physics-informed. The simulations obey incompressible Navier–Stokes,

$$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$6

but the training loss does not include $$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$7 or other PDE terms. Physical consistency is assessed a posteriori via a divergence metric

$$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$8

and the diffusion model is reported to have consistently lower $$\mathcal{G}^\star: \mathcal{C} \to \bigsqcup_{\Omega\in\Lambda} \mathcal{P}_2(\mathcal{U}(\Omega)), \quad \mathcal{G}^\star(\mathbf{c}) = p^\ast(\cdot\mid \mathbf{c}),$$9 than the CNN baseline across all geometries (Hu et al., 2024).

GeoFunFlow-3D introduces explicit physics guidance at several levels (Jiang et al., 25 Apr 2026). In latent space, it learns a time-dependent vector field under flow matching,

$$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$0

with a linear interpolation path $$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$1. The paper relates this choice to dynamic optimal transport and the Benamou–Brenier formulation, arguing that the resulting paths are approximate $$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$2 geodesics with zero material acceleration. It also states a global generation error bound

$$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$3

where $$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$4 is the intrinsic dimension of the physical manifold and $$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$5 is a Lipschitz constant of the ideal vector field (Jiang et al., 25 Apr 2026).

Physics enters GeoFunFlow-3D through a composite loss

$$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$6

with time-dependent homotopy weights. For external flows, $$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$7 incorporates a Navier–Stokes residual under a phase-field mask. For internal compressible flows, it includes the ideal gas law,

$$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$8

a shock mask

$$\text{geometry} \;\longmapsto\; \mathbf{z}_0(x,y) = [u(x,y), v(x,y), p(x,y), \omega(x,y)],$$9

and a thermodynamic loss

$u,v$0

The paper presents this as a mechanism for balancing data manifold learning and rigid physics constraints during training (Jiang et al., 25 Apr 2026).

Another notable component is the high-order discrete engine without automatic differentiation (No-AD). Derivatives on the latent grid are computed by a fourth-order central difference stencil,

$u,v$1

with one-sided five-point schemes on boundaries. The paper argues that, unlike automatic differentiation, the discrete operator has a band-limited Fourier symbol and therefore bounded NTK eigenvalues in the high-frequency regime (Jiang et al., 25 Apr 2026). This suggests that GeoFunFlow-3D is not merely geometry-conditioned generation but an attempt to reconcile generative transport, PDE residuals, and spectral stability within a single operator-learning framework.

6. Benchmarks, empirical behavior, and limitations

The 2025 GeoFunFlow paper evaluates on five inverse benchmarks: Darcy, Cylinder, Plasticity, Airfoil, and Ahmed body (Wang et al., 28 Sep 2025). The reported relative $u,v$2 test errors are summarized below.

Model	Darcy	Cylinder	Plasticity
Geo-FNO	0.0065	0.1298	0.0326
Transolver	0.0253	0.0993	0.0172
GeoFAE	0.0064	0.0538	0.0132
GeoFunFlow	0.0085	0.0567	0.0136

Model	Airfoil	Ahmed Body
Geo-FNO	0.1094	0.2272
Transolver	0.0641	0.0876
GeoFAE	0.0083	0.0820
GeoFunFlow	0.0087	0.0811

The paper notes that GeoFAE alone often attains the lowest reconstruction error, while GeoFunFlow stays close and remains significantly better than prior operator-learning baselines on complex geometries. It also reports that GeoFunFlow degrades more slowly under increased noise or reduced sampling fraction, and remains usable even at settings such as 5% sampling and 10% noise (Wang et al., 28 Sep 2025).

GeoFunFlow-3D evaluates on BlendedNet and NASA Rotor37 (Jiang et al., 25 Apr 2026). On BlendedNet with 100 training samples, the reported MAEs are $u,v$3 for $u,v$4, $u,v$5 for $u,v$6, and $u,v$7 for $u,v$8. On the Rotor37 blind 200-sample test, the reported RRMSEs are pressure $u,v$9, density $p$0, and temperature $p$1, while the 1000-sample limit yields pressure $p$2, density $p$3, and temperature $p$4 (Jiang et al., 25 Apr 2026). The abstract highlights the pressure-field RRMSE of $p$5 and states that the framework accurately captures 3D detached shock structures.

The 2D geometry-to-flow precursor reports superiority over CNN-based models and a VAE model on instantaneous flow prediction and on complex geometries such as a cross and “PKU,” with lower normalized $p$6 velocity error and smaller divergence across almost all geometries (Hu et al., 2024). In qualitative terms, the diffusion model is described as capturing asymmetric von Kármán vortex streets, boundary layers, separation patterns, and plausible wakes for strongly out-of-distribution shapes.

The literature also marks clear limitations. GeoFunFlow is currently described as benchmark-scale, and scaling to industrial-scale problems with millions of mesh elements remains challenging (Wang et al., 28 Sep 2025). It lacks explicit physics constraints in its current inverse formulation and is not explicitly equivariant to translations or rotations. GeoFunFlow-3D, although physics-guided, is still limited to steady RANS-based fields and remains vulnerable to extreme geometric distortions or out-of-distribution geometries, for which some fine-tuning may be required (Jiang et al., 25 Apr 2026). The 2D geometry-to-flow precursor is limited to 2D incompressible flow, static single-obstacle settings, implicit time handling, and fixed outer-domain boundary conditions (Hu et al., 2024).

Taken together, these works define GeoFunFlow as a geometry-driven generative operator-learning program with three progressively richer interpretations: geometry-to-flow diffusion on fixed CFD grids, latent posterior sampling for inverse problems on irregular geometries, and physics-guided flow matching for high-fidelity 3D aerodynamic inference (Hu et al., 2024, Wang et al., 28 Sep 2025, Jiang et al., 25 Apr 2026). The common thread is the replacement of deterministic geometry-to-field regression by geometry-conditioned generative modeling, with the aim of representing multi-modal field structure, handling complex geometries, and, in later variants, integrating explicit physical constraints.