WorksPhysics and Astronomy
Flow Mesh: A Velocity-Regularized Flow Matching Framework for Stable Distributional Modeling of Physical Fields on Unstructured Meshes
Abstract
Thermal-fluid field distributions underpin the analysis and operation of energy infrastructure, from nuclear-reactor cooling channels to industrial heat exchangers, but resolving these distributions through computational fluid dynamics requires prohibitively large simulation ensembles. Their complex geometries also demand unstructured meshes that are poorly accommodated by generative models developed for regular grids. Here we introduce FlowMesh, a generative framework that combines permutation-invariant mesh encoding with deterministic latent flow matching and an l1 penalty on the learned transport velocity. FlowMesh generates full-field realizations directly on fixed unstructured meshes without auxiliary grid interpolation or a stochastic denoising chain. Across lid-driven-cavity, pressurized-water-reactor subchannel and heat-exchanger datasets, evaluated over six matched training seeds, FlowMesh achieves macro-averaged probability-density-function coverage of 91.7%-97.0%, exceeding an architecture-controlled latent diffusion model by – percentage points. Latent-velocity regularization increases coverage by percentage points for reactor flows and 1.1 points for heat-exchanger flows, while reducing temperature-field Kullback–Leibler divergence by 1.7-2.4-fold. Each full-field realization requires 1.23 \pm 0.15 s, providing a 585–1902-fold wall-clock acceleration over computational fluid dynamics. FlowMesh therefore makes large-scale generative characterization of thermal-fluid states computationally feasible, providing a foundation for data-driven analysis and digital-twin development in fixed-geometry energy systems.
Introduction
Many decisions in physical and energy systems depend not on a single best-estimate field but on a distribution of plausible states. Nuclear thermal-hydraulic analysis, heat-exchanger design, virtual sensing, and uncertainty-aware digital twins require ensembles of velocity, temperature, pressure, and turbulence fields over uncertain operating conditions. High-fidelity finite-element and computational-fluid-dynamics solvers [1, 2] provide these fields by resolving the governing conservation equations [3], but generating sufficiently large ensembles remains computationally expensive [4]. This creates a fundamental scale mismatch between the cost of high-fidelity simulation and the number of realizations required for uncertainty propagation, probabilistic analysis, and near-real-time state estimation [5, 6].
A second difficulty is geometric. Engineering simulations are commonly defined on unstructured meshes whose nonuniform resolution conforms to curved boundaries, narrow flow passages, wall-adjacent gradients, and locally refined regions [7, 8]. In reactor subchannels and heat-transfer components, these geometric features are part of the physical problem rather than a numerical inconvenience [9, 10]. Mapping such fields to auxiliary Cartesian grids can introduce interpolation error and alter the representation of precisely the regions in which gradients are largest [11, 12]. A useful scientific generative model should therefore operate on the computational representation on which the high-fidelity data are produced.
Diffusion models provide a powerful route to high-dimensional generative modeling, but their standard formulation introduces an additional computational question. Samples are generated through an iterative reverse process that reconstructs data from progressively corrupted states [13, 14]. This strategy is highly successful in image generation, yet scientific field generation raises a different requirement: whether comparable or better distributional fidelity can be obtained through a continuous deterministic transport map while avoiding a long denoising chain. The distinction is especially relevant on learned latent representations of irregular meshes, where the geometry encountered by the generative model is induced by the encoder rather than supplied explicitly to the sampling process.
Iterative denoising can accumulate approximation and discretization error, and accelerated samplers may move intermediate states into regions in which the learned denoiser is less accurate [15, 16]. These observations do not imply that stochastic generative models are intrinsically incompatible with physical systems. Rather, they motivate a testable question: when the target data are high-dimensional physical fields represented on an irregular mesh, does deterministic transport through the learned latent space provide a more effective generative mechanism than iterative stochastic denoising? We address this question with an architecture-controlled comparison in which FlowMesh and a latent DDPM share the same PointNet autoencoder and latent training data. This design separates the representation problem from the generative-transport problem and allows the effect of the sampling mechanism to be assessed directly on the present benchmarks.

Figure 1. FlowMesh formulates mesh-based field generation as deterministic latent transport. A permutation-invariant PointNet autoencoder maps physical fields and node coordinates from a fixed unstructured mesh into a compact latent representation. A learned time-dependent velocity field then transports samples from a Gaussian source distribution to the latent data distribution through numerical integration of an ordinary differential equation. The decoder maps the terminal latent state back to the original mesh without interpolation to an auxiliary Cartesian grid. The schematic contrasts this continuous transport formulation with iterative denoising; it is conceptual and does not imply that individual learned trajectories coincide with a physical trajectory or an optimal-transport geodesic.
Here, we introduce FlowMesh, a generative framework for learning distributions of physical fields on fixed unstructured geometries through deterministic latent transport. Three design choices define the approach. First, a permutation-invariant PointNet autoencoder [17] jointly embeds node coordinates and node-associated physical quantities into a fixed-dimensional latent representation without interpolation onto an auxiliary grid. Second, flow matching [18] learns a continuous time-dependent velocity field whose ordinary differential equation transports samples from a Gaussian reference distribution to the latent distribution of the training fields. Third, an penalty on the latent velocity magnitude provides a tunable bias against unnecessarily large transport velocities. This penalty is not a Dirichlet-energy term and does not impose a formal Lipschitz constraint. Instead, its effect is evaluated empirically through a five-order-of-magnitude sweep in , together with measurements of distributional accuracy and the sampled Jacobian norm of the learned velocity field. The resulting experiments allow us to distinguish where velocity regularization materially improves latent transport from regimes in which it is unnecessary or detrimental.
FlowMesh addresses a different objective from the deterministic neural operators commonly used for forward surrogate modeling, including DeepONet, MIONet, S-DeepONet, the Fourier neural operator, and transformer-based operators [19, 20, 21, 22, 23]. In their standard supervised form, these models learn a mapping from prescribed inputs to a point prediction of the corresponding field. FlowMesh instead targets the distribution of full fields represented in a simulation ensemble. In the formulation studied here, this distribution is learned over the sampled operating envelope rather than conditioned on a user-specified operating point. The resulting samples are therefore suited to ensemble characterization and uncertainty-propagation tasks rather than deterministic forward prediction. Probabilistic neural operators and conditional generative operators represent an important adjacent class of methods; direct comparison with those formulations is outside the present scope. We therefore treat FlowMesh as complementary to deterministic operator learners rather than as a replacement for them.
We evaluate FlowMesh on three benchmarks that progressively increase the geometric and thermal complexity of the generated fields: a lid-driven cavity, a pressurized-water-reactor subchannel, and a heat exchanger. Six matched training seeds establish that velocity regularization is effectively neutral on the already accurate cavity benchmark but improves the mean distributional fidelity of the PWR and heat-exchanger ensembles, with the largest gain occurring for temperature. An autoencoder-only diagnostic reaches more than 99.4% PDF coverage in the PWR case, showing that the dominant remaining error arises in latent distribution transport rather than mesh reconstruction. Against a latent DDPM using the same PointNet representation, FlowMesh improves macro-averaged PDF coverage by 11.6–45.4 percentage points and reduces sampling cost. Relative to the reference CFD calculations, each generated realization is 585–1,902 times faster. At the same time, a no-slip diagnostic reveals a substantial boundary-condition residual, establishing an important boundary of the present result: high distributional fidelity does not by itself guarantee pointwise satisfaction of the governing constraints. Together, the experiments identify deterministic latent transport as a viable generative representation for physical-field ensembles on fixed unstructured meshes while showing that explicit physical constraints remain necessary for safety-critical deployment.
Results
Choice of benchmarks The three benchmarks are selected to test distinct aspects of the generative problem rather than to reproduce industrial scale. The lid-driven cavity provides a comparatively simple recirculating flow in which unregularized flow matching is already accurate and therefore acts as a useful negative control for the regularizer. The PWR subchannel introduces nuclear-relevant rod geometry, heated boundaries, and coupled velocity, turbulence, and temperature fields. The heat exchanger tests whether the observed behavior persists on a second energy-system geometry with a larger set of thermofluid variables. Each dataset retains the native fixed unstructured node set used by the CFD calculation. The objective is therefore narrower than universal mesh generalization: we test whether deterministic latent transport can reproduce distributions of high-dimensional fields on fixed irregular geometries, and whether velocity regularization provides a reproducible benefit when unregularized transport is less accurate.
Lid-driven cavity flow
We evaluate FlowMesh on the canonical lid-driven cavity (LDC) benchmark to assess its generative performance on an advection-dominated flow. Consider a two-dimensional lid-driven cavity defined on . A prescribed time-dependent velocity drives the upper boundary in the horizontal direction, while zero-velocity conditions are imposed on the remaining walls. The flow fields are obtained from the incompressible Reynolds-averaged Navier–Stokes equations with a – turbulence closure. Denoting the velocity by and the pressure by , the equations of continuity and momentum are:
where is the kinematic viscosity and the eddy viscosity.
The turbulent kinetic energy (TKE) field evolves as
where is the production term, the dissipation rate, and a model constant.
| Pressure | Velocity | TKE | |
| 0 | |||
| 1.0 |
Table 1. LDC velocity-regularization sweep across 6 independent training seeds (PDF coverage, mean standard deviation, evaluated against autoencoder-reconstructed reference fields). Coverage varies by less than approximately one percentage point for , whereas produces a large loss of fidelity in all three fields. The result establishes a broad low-to-moderate regularization regime in which the cavity solution is insensitive to , followed by clear over-regularization at the largest weight.
Boundary conditions are given by
A uniform pressure field and a zero velocity was set as a initial condition for this system:
We generate the lid velocity randomly and discretize it into 90 uniform time steps, serving as the input forcing profile. Simulations are performed in ANSYS FLUENT using an unstructured mesh with 4225 spatial nodes. The numerical solver outputs the velocity magnitude , the pressure field , and the TKE field at every timestep. At the terminal simulation time, these fields constitute the training dataset. We trained FlowMesh on 20,000 samples drawn from the full simulation dataset. Generated and reference fields are compared in Figure 2.

Figure 2. Lid-driven cavity benchmark results on an unstructured mesh. (a) Two-dimensional lid-driven cavity configuration with prescribed boundary conditions and the turbulent kinetic energy (TKE) field at the terminal simulation time. (b) Probability density function (PDF) comparisons between the reference computational fluid dynamics (CFD) data and FlowMesh-generated samples for pressure, velocity magnitude, and TKE. (c) Node-wise diagnostics for velocity magnitude, including mean and variance fields from the reference data and generated samples with and without regularization, mean-error maps relative to the reference, and Kullback-Leibler (KL) and Jensen-Shannon (JS) divergence maps between the reference and generated node-wise distributions.
Without regularization, FlowMesh attains PDF coverages of for TKE, for velocity magnitude, and for pressure, with corresponding KL divergences of , , and , respectively (Table 2). Node-wise statistics (Figure 2c) show that the mean and variance are reproduced across most of the domain, with the largest discrepancies concentrated near the moving lid and the bottom corners, where the flow gradients are strongest.
| Model | Field | KL ↓ | JS ↓ | PDF Coverage (%) ↑ |
| FlowMesh No Reg. | Pressure | 0.0071 ± 0.0058 | 0.0027 ± 0.0024 | 96.0 ± 1.6 |
| Velocity | 0.0173 ± 0.0063 | 0.0054 ± 0.0017 | 95.7 ± 1.0 | |
| TKE | 0.0027 ± 0.0019 | 0.0010 ± 0.0007 | 97.7 ± 1.3 | |
| FlowMesh , Reg. () | Pressure | 0.0060 ± 0.0041 | 0.0023 ± 0.0016 | 96.2 ± 1.2 |
| Velocity | 0.0180 ± 0.0071 | 0.0057 ± 0.0020 | 95.7 ± 0.8 | |
| TKE | 0.0023 ± 0.0018 | 0.0008 ± 0.0006 | 97.8 ± 1.2 | |
| FlowMesh , Reg. () | Pressure | 0.0057 ± 0.0030 | 0.0018 ± 0.0009 | 96.8 ± 0.9 |
| Velocity | 0.0178 ± 0.0074 | 0.0055 ± 0.0019 | 96.0 ± 0.7 | |
| TKE | 0.0017 ± 0.0009 | 0.0006 ± 0.0003 | 98.1 ± 0.8 |
Table 2. LDC distributional metrics over 6 training seeds (mean ± standard deviation). All configurations share the same architecture, data, and 100-epoch training schedule. Coverage is evaluated against autoencoder-reconstructed reference fields. Regularization changes coverage by less than one percentage point, within the seed-to-seed variability. Lower KL and JS and higher PDF coverage are better.
At , regularization increases the mean PDF coverage by only 0.8, 0.3, and 0.4 percentage points for pressure, velocity, and TKE, respectively. These changes are smaller than the corresponding seed-to-seed standard deviations, and the reported mean-±-standard-deviation intervals overlap for all three fields. The divergence metrics are also field dependent: KL and JS decrease modestly for pressure and TKE but change slightly unfavorably for velocity. The results therefore do not establish a clear LDC improvement from regularization. The summary statistics in Table 3 provide a complementary assessment. The unregularized model is closer to the original data in eight of the nine reported mean, minimum, and maximum statistics, whereas the regularized model is closer only for the maximum velocity. This does not contradict the marginal increase in PDF coverage, because coverage measures overlap of the full empirical distribution, while the reported means and extrema probe only selected statistics and are particularly sensitive to tail samples.
| Data Origin | Field | Mean | Min Value | Max Value |
| Original | Pressure | −234.7000 | −1570.1500 | 2515.7000 |
| Velocity | 0.0320 | 0.0000 | 0.7470 | |
| TKE | 0.3130 | 0.0000 | 7.4450 | |
| FlowMesh No Regularizer | Pressure | −226.9062 | −1239.5104 | 2140.1989 |
| Velocity | 0.0206 | −0.0051 | 0.7677 | |
| TKE | 0.3086 | −0.1371 | 6.9881 | |
| FlowMesh With Regularizer () | Pressure | −207.8638 | −1179.5140 | 2037.2308 |
| Velocity | 0.0203 | −0.0054 | 0.7500 | |
| TKE | 0.2966 | −0.1503 | 6.7631 |
Table 3. LDC (lid-driven cavity): comparison of statistical values of data across various fields. FlowMesh rows are the mean over 6 independent training seeds (same architecture, data and 100-epoch schedule; only the training seed differs); the Original row is as reported for the original dataset. Negative velocity-magnitude & TKE values are physically inadmissible; the table reports raw model outputs prior to any non-negativity post-processing.
Taken together, the LDC benchmark provides an important negative control. Unregularized flow matching already attains high distributional fidelity, and the changes produced by are smaller than the seed-to-seed variability for the principal coverage metrics. The data therefore do not support a claim that regularization universally improves FlowMesh. Instead, they show that a broad interval of low-to-moderate regularization leaves an already accurate transport model largely unchanged. At , coverage collapses to 50.8%, 14.9%, and 60.1% for pressure, velocity, and TKE, respectively, demonstrating that excessive suppression of the latent velocity can remove the transport capacity required to reach the target distribution.
PWR subchannel
We next test FlowMesh in a nuclear thermal-hydraulics setting in which distributional field modeling has direct engineering relevance. A pressurized-water-reactor (PWR) subchannel confines coolant between heated fuel rods, producing strongly nonuniform velocity, turbulence, and temperature fields on an irregular cross-section. Ensembles of such fields are relevant to uncertainty propagation, virtual sensing, and probabilistic characterization of thermal margins, but obtaining them through repeated CFD calculations is computationally expensive. The benchmark therefore tests whether latent generative transport remains accurate when the field distribution contains both coherent coolant motion and steep wall-adjacent thermal gradients. The operating variables comprise an axially varying wall heat-flux profile,
where is the heated length and is the heat-flux amplitude, together with the inlet temperature and inlet velocity magnitude . The resulting input space is
where and .
The response is recorded on a transverse plane at a fixed axial location . At every mesh node, the model generates the velocity magnitude , temperature , and turbulent kinetic energy . The output space therefore, is three-channel:
without interpolation of the unstructured field onto a Cartesian grid. We generated 5,000 reference simulations in ANSYS FLUENT on a three-dimensional unstructured mesh and extracted a common transverse section containing 1,733 nodes. The operating parameters were sampled independently as
Each condition thus maps to an array in . This fixed but geometrically irregular node set isolates the distribution-learning problem from mesh interpolation error while retaining the rod-wall boundary layers that make the case challenging.
Across six fixed, matched training seeds, unregularized FlowMesh attains PDF coverages of 89.6 5.0% for velocity, 81.6 6.6% for temperature, and 92.5 3.8% for TKE (Table 4). At , the corresponding means increase to 92.5 2.5%, 83.7 12.1%, and 95.1 1.9%. The gain is most reproducible for velocity and TKE, for which the across-seed variability also decreases. Temperature is more sensitive to training seed at this weight. Increasing the regularization strength to reduces the mean temperature KL divergence from 0.1517 to 0.0635 and increases temperature PDF coverage from 81.6% to 87.9%. This setting also gives the largest macro-averaged coverage, increasing it from 87.90% without regularization to 91.67%. The six-seed experiment therefore supports a field-dependent effect of velocity regularization: moderate regularization improves the mean distributional fidelity of all three PWR fields, but the weight that is most favorable differs across channels and the temperature response is less seed-stable than the velocity and TKE responses.
Table 4: PWR sub-channel distributional metrics over six fixed, matched training seeds (mean standard deviation). All configurations share the same architecture, data, and 100-epoch training schedule. FlowMesh is evaluated against autoencoder-reconstructed reference fields. Lower KL and JS and higher PDF coverage are better.
| Model | Field | KL | JS | PDF Coverage (%) |
| FlowMesh No Reg. | Velocity | 0.0613 0.0454 | 0.0221 0.0169 | 89.6 5.0 |
| Temperature | 0.1517 0.1041 | 0.0545 0.0339 | 81.6 6.6 | |
| TKE | 0.0350 0.0323 | 0.0118 0.0105 | 92.5 3.8 | |
| FlowMesh , Reg. () | Velocity | 0.0323 0.0186 | 0.0114 0.0055 | 92.5 2.5 |
| Temperature | 0.1350 0.1675 | 0.0536 0.0657 | 83.7 12.1 | |
| TKE | 0.0142 0.0095 | 0.0050 0.0031 | 95.1 1.9 | |
| FlowMesh , Reg. () | Velocity | 0.0397 0.0352 | 0.0126 0.0100 | 92.4 3.3 |
| Temperature | 0.0635 0.0286 | 0.0242 0.0103 | 87.9 4.0 | |
| TKE | 0.0188 0.0179 | 0.0063 0.0057 | 94.7 2.5 |
Table 4. PWR sub-channel distributional metrics over six fixed, matched training seeds (mean standard deviation). All configurations share the same architecture, data, and 100-epoch training schedule. FlowMesh is evaluated against autoencoder-reconstructed reference fields. Lower KL and JS and higher PDF coverage are better.
The field statistics in Table 5 provide a complementary view of these distributional improvements. The model is closer to the original data in five of the nine reported statistics, primarily the field extrema, whereas the unregularized model is closer for all three field means and the TKE minimum. Thus, the improvements in KL divergence, JS divergence, and PDF coverage do not translate into uniformly improved means and extrema.
Table 5: PWR sub-channel field statistics. Each FlowMesh entry is the average of the corresponding statistic over the 6 training seeds. The Original row reports statistics from the reference dataset. Negative velocity-magnitude & TKE values are physically inadmissible; the table reports raw model outputs prior to any non-negativity post-processing.
| Data Origin | Field | Mean | Min Value | Max Value |
| Original | Velocity | 3.992 | 0.000 | 5.250 |
| Original | Temperature | 573.880 | 512.260 | 649.680 |
| Original | TKE | 0.056 | 0.011 | 0.146 |
| FlowMesh, No Reg. | Velocity | 4.024 | 5.461 | |
| FlowMesh, No Reg. | Temperature | 575.795 | 503.519 | 659.990 |
| FlowMesh, No Reg. | TKE | 0.057 | 0.009 | 0.156 |
| FlowMesh, With Reg. () | Velocity | 3.931 | 5.415 | |
| FlowMesh, With Reg. () | Temperature | 576.075 | 503.580 | 658.834 |
| FlowMesh, With Reg. () | TKE | 0.054 | 0.008 | 0.153 |
Table 5.
The spatial diagnostics in Figure 3 complement these marginal metrics. The mean fields are reproduced across most of the coolant domain, while the largest KL and JS discrepancies remain concentrated near the heated rod surfaces, where the temperature and velocity gradients are steepest. This localization is consistent with temperature remaining the most challenging field, even though stronger regularization substantially improves its aggregate distributional metrics.

Figure 3. Pressurized water reactor (PWR) subchannel benchmark results on an unstructured mesh. (a) Schematic of the reactor subchannel geometry showing the fuel rods, coolant inlet conditions, flow direction, and extracted transverse cross-section, illustrated using the temperature field. (b) Probability density function (PDF) comparisons between the reference computational fluid dynamics (CFD) data and FlowMesh-generated samples for velocity magnitude, temperature, and turbulent kinetic energy (TKE), with and without regularization. (c) Node-wise diagnostics for temperature, including mean and variance fields from the reference data and generated samples with and without regularization, mean-error maps relative to the reference, and Kullback-Leibler (KL) and Jensen-Shannon (JS) divergence maps between the reference and generated node-wise distributions.
The separate single-run comparison in Table 6 attains its highest joint PDF coverage with supervision and . This configuration achieves coverages of 94.40% for velocity, 93.07% for temperature, and 96.98% for TKE, corresponding to a macro-average of 94.82%, compared with 94.31% for the regularized configuration. Relative to the unregularized baseline, regularization increases coverage by 1.47, 0.30, and 1.71 percentage points for velocity, temperature, and TKE, respectively.
Table 6: PWR sub-channel single-run comparison of and flow-matching supervision and the autoencoder reconstruction ceiling. For each loss, the unregularized baseline () is compared with the selected regularized setting (). The autoencoder reconstruction diagnostic passes reference fields through and evaluates them using the same metrics; it upper-bounds the distributional fidelity attainable by the latent flow model with this autoencoder. Lower KL, JS, and Wasserstein distance and higher PDF coverage are better.
| Model | Field | KL | JS | Wasserstein | PDF Coverage (%) |
| Flow Match – No Reg (, ) | Velocity | 0.0276 | 0.0112 | 0.0650 | 92.93 |
| Flow Match – No Reg (, ) | Temperature | 0.0322 | 0.0137 | 1.8165 | 92.77 |
| Flow Match – No Reg (, ) | TKE | 0.0117 | 0.0044 | 0.0016 | 95.27 |
| Flow Match – With Reg (, ) | Velocity | 0.0175 | 0.0075 | 0.0255 | 94.40 |
| Flow Match – With Reg (, ) | Temperature | 0.0310 | 0.0134 | 2.6921 | 93.07 |
| Flow Match – With Reg (, ) | TKE | 0.0047 | 0.0018 | 0.0006 | 96.98 |
| Flow Match – No Reg (, ) | Velocity | 0.0977 | 0.0311 | 0.1433 | 85.10 |
| Flow Match – No Reg (, ) | Temperature | 0.0275 | 0.0105 | 1.3739 | 92.17 |
| Flow Match – No Reg (, ) | TKE | 0.0469 | 0.0157 | 0.0036 | 89.77 |
| Flow Match – With Reg (, ) | Velocity | 0.0137 | 0.0053 | 0.0162 | 94.30 |
| Flow Match – With Reg (, ) | Temperature | 0.0282 | 0.0109 | 2.0575 | 92.39 |
| Flow Match – With Reg (, ) | TKE | 0.0054 | 0.0020 | 0.0003 | 96.25 |
| AE reconstruction (diagnostic) | Velocity | 0.0044 | 99.50 | ||
| AE reconstruction (diagnostic) | Temperature | 0.0400 | 99.60 | ||
| AE reconstruction (diagnostic) | TKE | 99.43 |
Table 6.
The temperature result is nevertheless metric-dependent. Under supervision, regularization slightly improves temperature coverage and KL divergence but increases the Wasserstein distance from 1.8165 to 2.6921. Under supervision, temperature coverage increases by only 0.22 percentage points, while KL, JS, and Wasserstein distance all increase slightly. The single-run comparison therefore supports regularization as the best joint-coverage configuration and shows clearer benefits for velocity and TKE than for temperature. The autoencoder reconstruction diagnostic attains more than 99.4% coverage and KL divergence below for every channel. Relative to this ceiling, the selected regularized configuration retains a temperature KL gap of approximately 0.0309, indicating that most of the remaining distributional error is introduced by the latent flow stage rather than by autoencoder reconstruction.
Marginal distributional agreement does not imply that each generated realization satisfies the governing constraints. We therefore examined two complementary diagnostics for the regularized model with . The global energy diagnostic is comparatively insensitive to the small negative TKE excursions: clipping these values to zero changes the system energy by less than 0.01%. The wall-boundary diagnostic exposes a more consequential limitation. Generated velocity at stationary walls has a mean residual of approximately 1.5 m/s, and forcing the wall values to zero after generation removes approximately 5% of the domain momentum and 3.5% of its kinetic energy. Distributional fidelity and physical admissibility are therefore distinct objectives in the present model. The latter cannot be recovered through benign post-processing and will require hard boundary constraints or constraint-preserving parameterizations in the decoder or latent dynamics.
Heat Exchanger

Figure 4. Heat exchanger benchmark results on an unstructured mesh. (a) Schematic of the heat exchanger configuration showing the flow direction, internal geometry, and transverse cross-section used for virtual sensing, with velocity magnitude visualized on the sensor plane. (b) Probability density function (PDF) comparisons between the reference computational fluid dynamics (CFD) data and FlowMesh-generated samples with and without regularization for turbulent kinetic energy (TKE), pressure, temperature, three velocity components, and velocity magnitude. (c) Node-wise diagnostics for velocity magnitude, including mean and variance fields from the reference data and generated samples with and without regularization, mean-difference maps relative to the reference, and Kullback–Leibler (KL) and Jensen–Shannon (JS) divergence maps between the reference and generated node-wise distributions.
We next test FlowMesh on a heat-exchanger geometry to determine whether the trends observed in the PWR subchannel persist in a second energy-system configuration with different wall geometry and a larger set of thermofluid variables. The wall heat flux follows the same axial profile used in the PWR benchmark,
where is the heat-flux amplitude. The profile is applied uniformly along the streamwise -direction,
and the operating envelope additionally varies the inlet temperature and inlet velocity magnitude . The parameter space is therefore
The response is evaluated on a vertical cross-section at fixed transverse location . The reported fields comprise turbulent kinetic energy, pressure, temperature, the three velocity components, and velocity magnitude:
These seven reported channels are generated jointly as separate outputs of the decoder, including , which is not computed from the generated components; they are the quantities used in the distributional comparisons of Table 7 and Table 8.
Table 7: Comparison of FlowMesh metrics (KL, JS, PDF coverage) on the Heat-Exchanger dataset, averaged over 6 independent training seeds. Coverage is evaluated against the autoencoder-reconstructed reference. Lower is better for KL and JS.
| Model | Field | KL ↓ | JS ↓ | PDF Coverage (%) ↑ |
| FlowMesh No Regularizer | TKE | 0.003 82 | 0.001 42 | 97.800 |
| FlowMesh No Regularizer | Pressure | 0.004 89 | 0.001 77 | 96.162 |
| FlowMesh No Regularizer | Temperature | 0.186 62 | 0.066 22 | 80.288 |
| FlowMesh No Regularizer | Z-Velocity | 0.000 38 | 0.000 14 | 99.196 |
| FlowMesh No Regularizer | Y-Velocity | 0.001 89 | 0.000 70 | 98.586 |
| FlowMesh No Regularizer | X-Velocity | 0.003 39 | 0.001 34 | 97.784 |
| FlowMesh No Regularizer | Velocity Magnitude | 0.003 21 | 0.001 25 | 97.914 |
| FlowMesh With Regularizer () | TKE | 0.001 78 | 0.000 63 | 98.426 |
| FlowMesh With Regularizer () | Pressure | 0.002 60 | 0.000 93 | 97.140 |
| FlowMesh With Regularizer () | Temperature | 0.111 42 | 0.044 06 | 84.487 |
| FlowMesh With Regularizer () | Z-Velocity | 0.000 21 | 0.000 08 | 99.393 |
| FlowMesh With Regularizer () | Y-Velocity | 0.000 83 | 0.000 29 | 98.980 |
| FlowMesh With Regularizer () | X-Velocity | 0.001 12 | 0.000 42 | 98.329 |
| FlowMesh With Regularizer () | Velocity Magnitude | 0.001 14 | 0.000 43 | 98.398 |
Table 7.
Table 8: Comparison of statistical values of data across various fields (Heat-Exchanger). FlowMesh rows are the mean over 6 independent training seeds; the Original row is as reported for the original dataset.
| Data Origin | Field | Mean | Min Value | Max Value |
| Original | TKE | 1.0780 | 0.0490 | 2.9885 |
| Original | Pressure | 6645.9000 | −13 974.3800 | 25 704.1000 |
| Original | Temperature | 292.3600 | 263.8500 | 322.5100 |
| Original | Z-Velocity | 0.0009 | −4.4910 | 4.5110 |
| Original | Y-Velocity | 0.1602 | −2.1800 | 3.4880 |
| Original | X-Velocity | 3.9170 | 0.0000 | 9.2520 |
| Original | Velocity Magnitude | 4.1920 | 0.0000 | 9.2730 |
| FlowMesh No Regularizer | TKE | 1.0800 | 0.0303 | 3.1224 |
| FlowMesh No Regularizer | Pressure | 6654.6775 | −14 611.2389 | 26 844.0859 |
| FlowMesh No Regularizer | Temperature | 293.0365 | 258.6028 | 328.3941 |
| FlowMesh No Regularizer | Z-Velocity | 0.0009 | −4.6226 | 4.6316 |
| FlowMesh No Regularizer | Y-Velocity | 0.1604 | −2.2456 | 3.5759 |
| FlowMesh No Regularizer | X-Velocity | 3.9197 | −0.1500 | 9.4946 |
| FlowMesh No Regularizer | Velocity Magnitude | 4.1951 | −0.1521 | 9.5210 |
| FlowMesh With Regularizer () | TKE | 1.0732 | 0.0384 | 3.0997 |
| FlowMesh With Regularizer () | Pressure | 6614.1274 | −14 477.3594 | 26 646.2246 |
| FlowMesh With Regularizer () | Temperature | 292.8551 | 258.5447 | 328.7100 |
| FlowMesh With Regularizer () | Z-Velocity | 0.0009 | −4.5966 | 4.6061 |
| FlowMesh With Regularizer () | Y-Velocity | 0.1599 | −2.2326 | 3.5566 |
| FlowMesh With Regularizer () | X-Velocity | 3.9070 | −0.1323 | 9.4524 |
| FlowMesh With Regularizer () | Velocity Magnitude | 4.1816 | −0.1296 | 9.4803 |
Table 8.
Table 7 shows that latent-velocity regularization improves the mean KL divergence, JS divergence, and PDF coverage for every reported heat-exchanger field. The macro-average PDF coverage increases from 95.390% without regularization to 96.450% at , corresponding to a gain of 1.060 percentage points. The largest improvement occurs for temperature, whose coverage increases from 80.288% to 84.487%, while its KL divergence decreases from 0.18662 to 0.11142. The remaining fields already attain coverage above 96% without regularization and consequently show smaller coverage gains, although their KL and JS divergences consistently decrease. These results indicate that regularization provides broad improvements across the heat-exchanger fields, with its principal benefit concentrated in the more challenging temperature distribution.

Figure 5. Conceptual interpretation of latent-velocity regularization. (a) Unregularized flow matching is free to use latent velocities of any magnitude required by the data-matching objective. Large or rapidly varying transport fields can increase sensitivity of the numerical trajectory to local approximation error. The illustration is schematic and should not be interpreted as a measured trajectory or direct evidence of “manifold detachment.” (b) FlowMesh augments the flow-matching loss with an penalty on the latent velocity magnitude. The regularization sweep in Figure 6 shows that increasing is empirically associated with a reduction in the sampled Jacobian norm of the learned velocity field, while moderate values can improve distributional fidelity. The implemented penalty is not a Dirichlet-energy objective and does not establish Wasserstein-geodesic optimality or formal contractivity.
The field statistics in Table 8 provide a complementary assessment. The regularized model is closer to the original data in 13 of the 21 reported mean, minimum, and maximum statistics, whereas the unregularized model is closer in seven and the two models tie in one. The regularized model improves most of the field extrema, while the unregularized model remains closer for several field means. Thus, the distributional improvements are broadly reflected in the summary statistics but are not uniform across every statistic. Because Table 7 and Table 8 report averages without seed-level standard deviations, these comparisons establish improvements in the mean metrics but do not by themselves demonstrate statistical significance across seeds.
Discussion
Deterministic transport versus latent diffusion Table 9 provides the most direct comparison between the two generative mechanisms considered here. FlowMesh achieves higher macro-averaged PDF coverage than the latent DDPM on all three benchmarks, with differences of approximately 11.6, 45.4, and 13.8 percentage points for the LDC, PWR subchannel, and heat exchanger, respectively. Because the two approaches use the same PointNet autoencoder and therefore operate on the same encoded data, the comparison controls for the principal geometric representation. It does not, however, constitute a universal comparison between flow matching and diffusion: the generative networks differ in parameterization and size, and the diffusion baseline uses a standard cosine-schedule DDPM with 250-step DDIM sampling rather than an exhaustively optimized scientific diffusion model. The experiments therefore establish that deterministic flow matching is substantially more effective than this architecture-controlled latent DDPM configuration on the present datasets.
| Dataset | FlowMesh (Ours) | Flow Matching (Lipman et al.) | Latent DDPM |
| Lid-driven Cavity | 96.97 | 96.47 | 85.39 |
| PWR Subchannel | 91.67 | 87.90 | 46.30 |
| Heat Exchanger | 96.45 | 95.39 | 82.70 |
Table 9. Comparison of PDF coverage (%) between FlowMesh, vanilla flow matching (Lipman et al.), and a latent DDPM baseline across benchmark datasets. FlowMesh and vanilla flow matching values are unweighted macro-averages across all reported fields using six fixed, matched training seeds. FlowMesh uses velocity regularization with , whereas vanilla flow matching uses the identical model with . All three models share the same PointNet autoencoder and latent space. Higher values indicate better distributional fidelity.
The result is consistent with two plausible advantages of FlowMesh. First, its sampling trajectory is obtained by integrating a learned velocity field rather than repeatedly estimating a reverse denoising update. Second, the adaptive ODE solver requires fewer network evaluations than the 250-step DDIM sampler used here. These mechanisms are consistent with the measured accuracy and cost differences but are not independently isolated by the present experiments. We therefore interpret the diffusion comparison as evidence for the effectiveness of deterministic latent transport on these mesh-based physical-field distributions, not as evidence that stochastic generative modeling is intrinsically incompatible with physical systems.
Computational cost: deterministic transport versus iterative denoising Table 10 compares the measured computational cost of FlowMesh and the latent DDPM. Under the common 100-epoch training budget and batch size of 128, the recorded training times are 7.6 min for FlowMesh and 39.5 min for DDPM. This wall-clock difference should not be interpreted as an algorithm-only speed-up because the reported implementations use different numerical precision, bf16 for FlowMesh and fp32 for DDPM, and different network sizes. The sampling comparison is more directly interpretable. FlowMesh requires adaptive velocity-network evaluations per generated sample, whereas the selected DDIM sampler uses 250 denoising evaluations. The resulting measured inference costs are approximately 1.4 TFLOP and 2.6 TFLOP per sample, respectively.
| Metric | FlowMesh | Latent DDPM (shared PointNet encoder) |
| Network parameters | 20.8 M | 6.1 M |
| Forward pass | 15.1 GFLOP | 10.6 GFLOP |
| Training wall-clock (100 epochs) | 7.6 min | 39.5 min |
| Peak training memory | 9.4 GB | 6.1 GB |
| NFE per sample | 94.7 11.6 (adaptive) | 250 (DDIM) |
| Inference cost per sample | 1.4 TFLOP | 2.6 TFLOP |
| Wall-clock per sample | 1.23 0.15 s | 2.24 0.003 s |
| Sampling memory (batch 1) | 200 MB | 310 MB |
| Hardware | A100 (40 GB) | A100 (40 GB) |
Table 10. Computational cost of FlowMesh versus the latent DDPM baseline. Both models operate on identical PointNet latent spaces; reported differences reflect only the generative component. Measured on one NVIDIA A100 (40 GB) GPU on the 4,096-dimensional latent (reshaped to 1 64 64), which is identical across datasets; training times were measured on the PWR subchannel latents, with FlowMesh at ; FLOPs from the PyTorch operator-level profiler. Training uses a matched 100-epoch budget and a batch size of 128; FlowMesh trains in mixed precision (bf16), and DDPM in fp32, following each reference implementation.
End-to-end, one FlowMesh realization, including latent ODE integration and decoder reconstruction, requires s on a single NVIDIA A100 GPU, compared with 2.24 s for the DDIM implementation. The larger practical gain is relative to high-fidelity simulation: depending on the benchmark, FlowMesh is 585–1,902 times faster than the corresponding ANSYS FLUENT calculation (Table 11). Thus, in the implementations evaluated here, deterministic latent transport provides both higher distributional fidelity and lower sampling cost than the latent DDPM while reducing the cost of generating additional thermofluid realizations by approximately three orders of magnitude relative to CFD.
| Dataset | Mesh Nodes | CFD (s/sample) | FlowMesh (s/sample) | Speed-up{sup}†` |
| Lid-driven Cavity (LDC) | 4225 | 2340 | 1.23(15) | 1902 |
| PWR Subchannel | 1733 | 720 | 1.23(15) | 585 |
| Heat Exchanger (HX) | 3977 | 1800 | 1.23(15) | 1463 |
Table 11. Per-sample computation time for each benchmark dataset. CFD reference times correspond to single-case ANSYS FLUENT runs on a 12-core Xeon CPU. FlowMesh inference is measured on a single NVIDIA A100 (40 GB) GPU (batch size 1, adaptive Dormand–Prince ODE solver, regularised model ), averaged over 1,000 samples following 20 warm-up runs. Speed-up is the ratio of the CFD wall-clock time to the FlowMesh wall-clock time.
Relation to neural operators The present work targets distributional generation rather than deterministic forward regression. Neural operators such as DeepONet [19], MIONet [20], S-DeepONet [24], the Fourier neural operator [22], geometry-informed operators [25], and transformer-based operators are commonly trained to predict a field associated with a specified input function or parameter set. Under standard pointwise regression losses, these models are naturally evaluated through prediction error at that operating condition. FlowMesh instead generates an ensemble of fields whose distribution approximates the distribution represented by the training simulations. Because each reference solution is a deterministic RANS field, this distribution is the push-forward of the operating-condition distribution used to generate the dataset, (Appendix § 8.3). It therefore characterizes the sampled operating envelope (uniform over the ranges of Eq. (10) for the PWR subchannel) rather than a physical operational-uncertainty distribution, and when and are known the same ensemble can also be obtained by evaluating a deterministic operator on sampled inputs; FlowMesh learns it from field samples alone.
The distinction is particularly relevant in energy applications in which an ensemble, rather than a single best-estimate field, is required. Examples include estimating the distribution of peak coolant temperature across an uncertain operating envelope, generating synthetic field ensembles for downstream uncertainty propagation, and producing candidate states for sensor-placement or probabilistic-analysis studies. The present implementation should not, however, be interpreted as a conditional neural operator: as formulated in the Methods, the latent velocity field is not explicitly conditioned on a requested operating point. Conditional generation at prescribed heat flux, inlet temperature, or inlet velocity is therefore a natural extension rather than a capability demonstrated here.
Probabilistic neural operators, stochastic operator models, ensembles, and Bayesian formulations also address uncertainty in function-space prediction and are scientifically relevant comparators. We do not claim that generative flow matching is the only mechanism capable of representing field distributions. The comparison in this work is deliberately narrower: given a common PointNet representation, we compare deterministic latent flow matching, its velocity-regularized variants, and a standard latent DDPM. Deterministic neural operators remain complementary when fast point prediction at specified operating conditions is the primary objective.
Decomposing reconstruction and transport error The two-stage architecture introduces two distinct sources of error: compression and reconstruction through the autoencoder, and distributional transport in latent space. The autoencoder-only diagnostic in Table 6 attains PDF coverage above 99.4% and KL divergence below for all three PWR fields. By contrast, the selected regularized flow model yields a temperature KL divergence of 0.0310 and PDF coverage of 93.07%. The approximately 0.0309 KL gap relative to the autoencoder reconstruction therefore arises predominantly after encoding, indicating that latent distribution transport, rather than mesh reconstruction, is the principal remaining source of error in this benchmark. We use the autoencoder result as an empirical reconstruction reference rather than as a formal mathematical upper bound on every possible decoded distribution.
Limitations Several limitations define the present scope. First, all experiments use a fixed mesh for each physical system. PointNet provides permutation invariance to node ordering, but the study does not demonstrate zero-shot transfer across mesh topologies, resolutions, or geometries. Second, the generative velocity field is not explicitly conditioned on operating parameters in the formulation reported here. FlowMesh therefore learns the field distribution over the sampled operating envelope rather than a conditional distribution at an arbitrary prescribed operating point. Extending the model to conditional transport is necessary for many digital-twin applications.
Third, distributional fidelity does not guarantee physical admissibility. The PWR model exhibits a mean no-slip wall residual of approximately 1.5 m/s, and post-hoc correction changes domain momentum and kinetic energy non-negligibly. Hard boundary conditions, conservation constraints, or constraint-preserving output parameterizations will therefore be required before the method can be considered for safety-critical operational use. Small negative values also occur in quantities such as velocity magnitude and TKE, showing that positivity is not enforced by construction.
Fourth, the regularization weight is problem dependent. The principal six-seed comparisons use , whereas the separate PWR loss ablation and computational-cost measurements use . The sweep demonstrates a broad useful regime but does not provide an automatic selection rule. Fifth, the latent DDPM shares the same PointNet representation and training data but is not parameter-matched to the FlowMesh velocity network and uses a standard 250-step DDIM sampler. The reported diffusion gap therefore applies to this controlled baseline rather than to the full class of diffusion or score-based generative models.
Finally, the present benchmarks are fixed-geometry thermofluid field ensembles. Generalization to transient three-dimensional systems, discontinuous multiphase or reacting flows, previously unseen geometries, and out-of-distribution operating regimes remains to be established. These limitations bound the current conclusion to deterministic generative transport of distributions on the fixed unstructured geometries evaluated here.
The present results establish a narrower foundation for those applications. Permutation-invariant encoding can compress fixed unstructured-mesh fields with little reconstruction loss, deterministic flow matching can reproduce their distributions more accurately and cheaply than the latent DDPM evaluated here, and moderate velocity regularization can improve transport in the benchmarks for which the unregularized model is less accurate. The next scientific step is not simply to scale the architecture, but to integrate conditional operating variables, hard conservation and boundary constraints, topology-changing geometries, and transient dynamics into the generative process. Demonstrating those capabilities would determine whether deterministic latent transport can progress from efficient ensemble generation toward trustworthy generative digital twins for nuclear and other safety-critical energy systems.
Conclusion
FlowMesh shows that deterministic latent transport can generate distributions of high-dimensional physical fields on fixed unstructured meshes without interpolation to an auxiliary grid. Across lid-driven-cavity, PWR-subchannel, and heat-exchanger benchmarks, it attains 91.7–97.0% macro-averaged PDF coverage and outperforms a latent DDPM built on the same PointNet representation by 11.6–45.4 percentage points, while reducing the cost of each additional realization by roughly three orders of magnitude relative to the reference CFD. The effect of velocity regularization is regime dependent: it leaves the already accurate cavity model essentially unchanged but improves mean coverage where unregularized transport is less accurate, most strongly for temperature. An autoencoder-only diagnostic locates the remaining error primarily in latent transport rather than in mesh reconstruction.
Nuclear thermal hydraulics provides both the most demanding test case and the clearest limitation. In the PWR subchannel, high distributional fidelity coexists with a mean no-slip residual of approximately 1.5 m/s at stationary walls, showing that matching field statistics does not by itself enforce physical admissibility. Integrating hard boundary and conservation constraints into the decoder or the latent dynamics is therefore the next scientific challenge, together with conditioning on operating variables and extending the approach to topology-changing geometries and transient dynamics.
With such constraints in place, inexpensive ensemble generation after a high-fidelity dataset has been constructed could support rapid uncertainty propagation across coolant states, synthetic field ensembles for sensor-placement studies, and probabilistic exploration of thermal margins in nuclear reactors, heat exchangers, and other energy systems in which complex geometry makes repeated CFD evaluation costly.
Methods
Overview and rationale for the coupling
FlowMesh separates geometric compression from distributional generation. A PointNet-based autoencoder first maps a field defined on a fixed unordered node set to a compact latent representation. Because the same node-wise transformation is followed by a symmetric pooling operation, the encoded representation is invariant to permutations of the input ordering. A flow-matching model is then trained on the encoded states and learns a deterministic time-dependent velocity field whose ordinary differential equation transports samples from a Gaussian reference distribution to the latent data distribution. The generative variable parameterizes this probability-transport path and is not physical simulation time.
The autoencoder and flow model are trained sequentially. This design allows reconstruction error and latent-transport error to be diagnosed separately, as demonstrated by the autoencoder-only experiment in Table 6. We additionally augment the flow objective with an penalty on the latent velocity magnitude. The penalty does not directly constrain derivatives of the velocity field and therefore is not a formal Lipschitz or Dirichlet-energy regularizer. Its influence on local velocity-field variation is assessed empirically through the Jacobian-norm diagnostic in Figure 6. The regularization sweep is used to determine whether the penalty improves generation, leaves it unchanged, or suppresses transport excessively.

Figure 6. Regularization sweep maps the accuracy–stability trade-off in latent transport. PWR subchannel benchmark, single-run sweep. (a) Temperature KL divergence as a function of regularization strength . Under supervision, KL decreases from the unregularized value of approximately 0.032 to approximately 0.022 at and remains comparatively low through moderate , before increasing sharply at . The model deteriorates at smaller regularization weights. (b) Empirical Jacobian-norm proxy estimated on the sampled evaluation set . The proxy decreases from approximately 65 at small to approximately 28 for and 32 for at . This is an empirical measure of local velocity-field variation, not a proof of a global Lipschitz bound. (c) Temperature and velocity PDF coverage versus . Small-to-moderate regularization improves both fields in this single run, whereas excessive regularization causes both to collapse. Together, the panels show that useful regularization occurs in an intermediate regime: reducing latent-velocity freedom can improve distributional transport, but suppressing it too strongly prevents the model from reaching the target distribution.
Operating in latent space also reduces the dimensionality presented to the generative model relative to the original mesh fields. The actual computational consequences of this compression are quantified through measured network evaluations, FLOPs, memory, and wall-clock sampling time in Table 10, rather than inferred from dimensionality alone.
Latent Encoding of Unstructured Mesh
Each sample is a field on a fixed mesh of nodes, , where are node coordinates and are the node-associated physical quantities ( for the cavity and subchannel, for the heat exchanger). Each field channel is standardized to zero mean and unit variance over all samples and nodes; coordinates are left unscaled. A separate autoencoder is trained for each dataset.
The encoder follows PointNet [17] but uses mean rather than max pooling. Each node input passes through a shared node-wise MLP with widths 64, 128 and 256 and ReLU activations,
and the node embeddings are averaged,
which makes the encoding invariant to node ordering. A linear layer then maps the pooled feature to the latent,
The reshape to a array has no spatial meaning; it allows both generative models to use standard image U-Net architectures. Because is an affine function of , the latent distribution lies on an affine subspace of dimension at most 256 within . The latent size is identical for all three datasets, compared with , and field values for the cavity, subchannel and heat exchanger, respectively.
The decoder maps back to the fields on the fixed node set. Two convolutional layers ( channels) are followed by adaptive average pooling to and fully connected layers whose output has size , so that each output entry corresponds to a fixed node of the mesh. Node coordinates are encoder inputs only and are not reconstructed; each trained decoder is therefore specific to its mesh.
The autoencoder minimizes the mean-squared reconstruction error of the standardized fields,
using Adam (learning rate , batch size 32, 50 epochs). Reconstruction fidelity is assessed by the relative mean-squared error of the de-standardized fields and by the distributional diagnostic in Table 6.
Model architecture
In our methodology, we have employed Flow Matching (FM), a simulation-free framework that learns a time-dependent velocity field whose flow transports samples from a simple source distribution to the target data distribution. FM constructs a continuous-time probability path interpolating between a Gaussian source distribution and the target latent-data distribution , and trains a neural velocity field to reproduce the instantaneous velocities that generate this path. The flow defined by is then used in sampling by integrating with the following ODE:
Model parametrization
The velocity field is parameterized by a U-Net of the ADM architecture [26] operating on the latent array, with 64 base channels, channel multipliers , two residual blocks per resolution, single-head self-attention at the feature resolution, and no dropout (20.8 M parameters). The generative time enters through sinusoidal embeddings. The network receives no other input: neither the operating parameters nor any observation of the field is provided, so the model learns the unconditional latent distribution . Its output is a velocity with the same shape as . The same configuration is used for all three datasets and is trained in bfloat16 mixed precision.
Training
Using a straight-line interpolation between independent samples and , the network is trained by minimizing either the or error between the predicted and target velocities:
or
Here,
Unless otherwise stated, results labeled use , whereas results labeled use . Following [27], “conditional” in refers to conditioning the target velocity on the endpoint , not on physical inputs.
Here, is the interpolated sample at time , and is the ground-truth instantaneous velocity. During each training iteration, batches of pairs are drawn, and an independent value is sampled uniformly in for each pair. The interpolated samples and their corresponding target velocities are computed analytically. The network parameters are optimized using the Adam optimizer (, ) with a constant learning rate.
Sampling
To generate novel samples from the learned model, we begin by drawing an initial point from the Gaussian distribution and evolving it forward in continuous time according to the learned probability flow . For numerical integration, we employ the adaptive Dormand-Prince method, which provides a favorable balance between accuracy and computational efficiency for this task. The solver adjusts its step size adaptively based on local error estimates, with relative and absolute tolerances tuned on a held-out validation set.
Latent diffusion baseline
To separate geometric representation from the choice of generative mechanism, the diffusion baseline uses the same trained PointNet autoencoder as FlowMesh, including the same encoder , decoder , latent representation, and reconstruction objective . The encoded training samples presented to the two generative models are therefore identical. The latent diffusion model is a denoising diffusion probabilistic model (DDPM) [13] trained with a cosine noise schedule and sampled with a 250-step DDIM procedure. The noise-prediction network is a U-Net operating on the same latent array, with encoder widths , bottleneck widths , 128-dimensional time embeddings, and four-head self-attention with group normalization in every block (6.1 M parameters), using diffusion steps.
The comparison controls the geometric representation, dataset, batch size, and evaluation pipeline, but it is not parameter matched: the FlowMesh and DDPM generative networks have different parameter counts and computational graphs, as reported in Table 10. Their training implementations also use different numerical precision. Consequently, differences in distributional accuracy isolate the generative pipelines under a shared latent representation, but should not be interpreted as a theorem that flow matching is intrinsically superior to all diffusion architectures.
Latent-velocity regularization and stability diagnostics
Flow matching learns a velocity field whose ODE
transports probability mass from the reference distribution toward the latent data distribution. Under the usual regularity assumptions, the corresponding density satisfies the continuity equation
We augment the flow-matching objective with an penalty on the magnitude of the latent transport velocity:
where controls the strength of the penalty. The regularizer directly discourages large latent velocities. It does not contain derivatives of and therefore should not be interpreted as a Dirichlet-energy penalty or as direct mathematical control of the Lipschitz constant. Unlike kinetic-energy and Jacobian-norm regularizers for continuous-time flows [28], which penalize and estimate , respectively, the penalty uses only the network output, adds negligible cost, and requires no stochastic trace estimator.
To determine whether reducing velocity magnitude is associated with a smoother learned vector field, we separately evaluate the Jacobian after training. We report the empirical proxy
where is the finite evaluation set used in the regularization sweep. A decrease in indicates lower local sensitivity over the sampled region, but does not constitute a global Lipschitz bound. Figure 6 shows that increasing is empirically associated with a decrease in this Jacobian-norm proxy. The same sweep also shows that this reduction is beneficial only over an intermediate range: excessive regularization suppresses the transport sufficiently strongly that distributional fidelity collapses.
Flow matching is closely related to continuous probability transport, and optimal transport provides useful mathematical context for such dynamics. The Benamou–Brenier formulation of the Wasserstein-2 distance minimizes an kinetic action of the form
subject to the continuity equation and prescribed endpoint distributions. The velocity penalty used in FlowMesh is a different objective and does not imply recovery of the Benamou–Brenier minimizer or a Wasserstein-2 geodesic. We therefore use optimal transport only as conceptual context for continuous distribution transport, not as a theoretical characterization of the learned FlowMesh trajectories.
Acknowledgments
This work used the Delta and DeltaAI systems at the National Center for Supercomputing Applications [awards OAC 2005572 and OAC 2320345] through allocation CIS240093 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.
Declaration of Generative AI and AI-Assisted Technologies
Generative AI tools were used during manuscript preparation for language editing, structural organization, and refinement of selected visual materials. All numerical experiments, reported quantitative results, scientific interpretations, references, and conclusions were independently checked and approved by the authors. The authors take full responsibility for the accuracy, integrity, and originality of the final manuscript.
Appendix
Random Vectors
Let be a probability space. A random vector in is a measurable function
such that for every Borel set , the event
The law (or distribution) of is the push-forward measure
If is absolutely continuous with respect to the Lebesgue measure on , then there exists a probability density function (pdf) such that
Expectation is defined for an integrable random vector as
Conditional density and Expectation
Let be a random vector on with joint density . The marginal densities are given by
If , then the conditional density of given is defined as
The conditional expectation of given is
More generally, conditional expectation is defined as a random variable satisfying:
is -measurable, where is the -algebra generated by .
For every bounded measurable ,
Diffeomorphisms and push-forward maps
Let and be smooth manifolds. A map is a diffeomorphism if:
is bijective,
is smooth,
is smooth.
If is a measure on , then the push-forward measure under is
If and , then . If admits density and is differentiable, then the density of is
Velocity fields and Flows
A vector field on a smooth manifold is a smooth section of the tangent bundle, that is, a mapping
such that for all .
Let be a time-dependent vector field. The associated flow is defined as the solution to the ordinary differential equation
If is measurable in and Lipschitz continuous in , then the flow exists and is unique, and each map is a diffeomorphism.
In flow matching, the vector field is parameterized by a neural network . The generative model is defined by the probability flow ordinary differential equation.
where is a prescribed reference distribution. In the idealized setting, the terminal random variable satisfies , where denotes the target data distribution.
Constructing probability paths
Let and be two probability measures on . A probability path is a family interpolating between them, i.e.
One natural choice is the linear interpolation in sample space:
This induces a probability path .
The instantaneous velocity of this interpolation is
Alternatively, one can define optimal transport paths. If is the optimal transport map sending to , then the McCann interpolation is
This path minimizes the quadratic Wasserstein distance and provides a geodesic in Wasserstein space.
Flow matching also constructs continuous probability paths, but a generic flow-matching path need not be an optimal-transport geodesic. In the present implementation, training uses straight-line interpolation between independently sampled source and target latent states. The learned velocity field is therefore trained to reproduce that prescribed probability path rather than to solve the Monge–Kantorovich or Benamou–Brenier optimization problem. The connection to optimal transport is conceptual unless the coupling and training objective are explicitly chosen to enforce optimal-transport structure.
References
- [1]Olgierd C Zienkiewicz and Robert Leroy Taylor. The finite element method set. Elsevier, 2005.
- [2]John David Anderson, John Wendt, et al. Computational fluid dynamics, volume 206. Springer, 1995.
- [3]Alfio Quarteroni and Alberto Valli. Numerical approximation of partial differential equations. Springer, 1994.
- [4]Peter Benner, Serkan Gugercin, and Karen Willcox. A survey of projection-based model reduction methods for parametric dynamical systems. SIAM review, 57(4):483–531, 2015.DOI
- [5]Kazuma Kobayashi and Syed Bahauddin Alam. Deep neural operator-driven real-time inference to enable digital twin solutions for nuclear energy systems. Scientific reports, 14(1):2101, 2024.DOI
- [6]Yuantao Yao, Te Han, Jie Yu, and Min Xie. Uncertainty-aware deep learning for reliable health monitoring in safety-critical energy systems. Energy, 291:130419, 2024.DOI
- [7]Pascal Jean Frey and Paul-Louis George. Mesh generation: application to finite elements. Iste, 2007.
- [8]Michael M. Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. Geometric deep learning: Going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017.DOI
- [9]Joe F Thompson. Grid generation techniques in computational fluid dynamics. AIAA journal, 22(11):1505–1523, 1984.DOI
- [10]Dimitri J Mavriplis. Directional agglomeration multigrid techniques for high-reynolds-number viscous flows. AIAA journal, 37(10):1222–1230, 1999.DOI
- [11]Joe F Thompson, Bharat K Soni, and Nigel P Weatherill. Handbook of grid generation. CRC press, 1998.
- [12]Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 8459–8468. PMLR, 13–18 Jul 2020.
- [13]Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020.arxiv.org/abs/2006.11239
- [14]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022.arxiv.org/abs/2112.10752
- [15]Wenpin Tang and Hanyang Zhao. Contractive diffusion probabilistic models. SIAM Journal on Imaging Sciences, 19(2):1364–1389, 2026.arxiv.org/abs/2401.13115
- [16]YAO Yuzhe, Jun Chen, Zeyi Huang, Haonan Lin, Mengmeng Wang, Guang Dai, and Jingdong Wang. Manifold constraint reduces operatorwise bias in accelerated diffusion sampling. In The Thirteenth International Conference on Learning Representations, 2025.
- [17]Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660, 2017.arxiv.org/abs/1612.00593
- [18]Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022.arxiv.org/abs/2210.02747
- [19]Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218–229, 2021.
- [20]Pengzhan Jin, Shuai Meng, and Lu Lu. Mionet: Learning multiple-input operators via tensor product. SIAM Journal on Scientific Computing, 44(6):A3490–A3514, 2022.arxiv.org/abs/2202.06137
- [21]Junyan He, Shashank Kushwaha, Jaewan Park, Seid Koric, Diab Abueidda, and Iwona Jasiuk. Sequential deep operator networks (s-deeponet) for predicting full-field solutions under time-dependent loads. Engineering Applications of Artificial Intelligence, 127:107258, 2024.DOI
- [22]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895, 2020.arxiv.org/abs/2010.08895
- [23]Zhongkai Hao, Zhengyi Wang, Hang Su, Chengyang Ying, Yinpeng Dong, Songming Liu, Ze Cheng, Jian Song, and Jun Zhu. Gnot: A general neural operator transformer for operator learning. In International conference on machine learning, pages 12556–12569. PMLR, 2023.arxiv.org/abs/2302.14376
- [24]Somdatta Goswami, Aniruddha Bora, Yue Yu, and George Em Karniadakis. Physics-informed deep neural operator networks, 2022.arxiv.org/abs/2207.05748
- [25]Zongyi Li, Nikola Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Otta, Mohammad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, et al. Geometry-informed neural operator for large-scale 3d pdes. Advances in Neural Information Processing Systems, 36:35836–35854, 2023.arxiv.org/abs/2309.00583
- [26]Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021.
- [27]Yaron Lipman, Marton Havasi, Peter Holderrieth, Neta Shaul, Matt Le, Brian Karrer, Ricky TQ Chen, David Lopez-Paz, Heli Ben-Hamu, and Itai Gat. Flow matching guide and code. arXiv preprint arXiv:2412.06264, 2024.arxiv.org/abs/2412.06264
- [28]Chris Finlay, Jörn-Henrik Jacobsen, Levon Nurbekyan, and Adam Oberman. How to train your neural ode: the world of jacobian and kinetic regularization. In International conference on machine learning, pages 3154–3164. PMLR, 2020.arxiv.org/abs/2002.02798