WorksPhysics and Astronomy
Physics-Guided Generative Surrogates for Parametric Rarefied Flows with Neural-Field Auto-Decoders: A Pipeline-Level Study of Flow Matching and Diffusion
Abstract
We present a conditional latent generative framework for parametric rarefied flows that separates neural-field representation, latent transport, and frozen physics adaptation. Neural-field auto-decoders compress discrete-velocity cavity solutions and direct simulation Monte Carlo cylinder solutions into shared coordinate decoders. Train-only principal-component charts support conditional flow matching (FM) and diffusion without a deterministic condition-to-latent backbone, and structured low-rank adapters correct selected decoder outputs while the upstream pipeline remains frozen. On two steady benchmarks, the frozen pipelines interpolate out-of-sample conditions with cavity kinetic relative errors at the level and cylinder per-field area-weighted RMSEs of 0.038 (density), 0.041 (temperature), and below 0.01 (velocities). For the cavity, physics adaptation reduces the matched-grid Bhatnagar–Gross–Krook diagnostic by 28.65% while preserving field accuracy; for the cylinder, the analytic wall map enforces no-penetration exactly and, jointly with the learned FM adapter, reduces the inlet violation to 0.277 and the global mass-balance ratio to 0.963 of the frozen values with negligible field-error change. A five-seed controlled comparison with deterministic condition-to-chart multilayer perceptrons shows that, although the generative pipelines do not surpass the compact MLP in point accuracy on these single-valued steady problems, the results validate sampling-based conditional transport on the shared representation as an effective steady surrogate, with a natural route to multivalued or stochastic solution families.
Introduction
Rarefied gas flows arise in high-altitude and hypersonic aerodynamics, atmospheric reentry, and micro- and nanoscale transport systems. Departure from local thermodynamic equilibrium is commonly characterized by the Knudsen number , where is the molecular mean free path and is a characteristic length. As increases through the slip, transition, and free-molecular regimes, the constitutive closure underlying the Navier–Stokes equations becomes progressively inadequate, and kinetic descriptions that resolve molecular velocity distributions become necessary. Established approaches include the direct simulation Monte Carlo method (DSMC)1, the discrete velocity method (DVM)2,3, high-order lattice Boltzmann formulations4–6, and multiscale gas-kinetic schemes7–10. These methods provide access to strongly nonequilibrium states but remain expensive in many-query studies: deterministic kinetic solvers discretize the coupled physical–velocity phase space, whereas DSMC requires extensive particle sampling and often local mesh refinement. Repeated simulations over , Mach number, and other physical or geometric controls therefore yield large, heterogeneous datasets for which storage, reconstruction, and rapid condition-to-field prediction are separate computational challenges.
This computational burden has motivated data-driven surrogates at several levels of the simulation hierarchy. In continuum fluid mechanics, machine learning has been used for field reconstruction from sparse observations11–13 and for learned components that accelerate conventional solvers14, within a broader movement toward data-assisted physical modeling15,16. For many-query partial differential equations, the deep operator network (DeepONet), the Fourier neural operator, and conditional neural fields approximate mappings from input functions or physical parameters to solution fields17–19. In kinetic and rarefied-gas modeling, related efforts have targeted reduced Bhatnagar–Gross–Krook (BGK) representations20, learned moment closures21–23, structure-preserving collision operators24,25, and field-level surrogates for microscale flows26,27. Existing studies show that machine learning can approximate selected solver components and flow observables. The present task, however, combines three distinct requirements: compressing each high-dimensional flow solution, predicting unseen physical conditions from a limited number of parameter cases, and retaining the governing and boundary constraints that can be evaluated from the available variables. Here, data scarcity occurs in the parameter space rather than necessarily in the spatial domain: high-fidelity solutions are typically available at only a limited number of sampled parameter combinations, whereas each sampled condition can be associated with multiple spatial fields and, when the kinetic state is represented directly, high-dimensional discrete velocity-distribution functions28,29. Accurate reconstruction of the observed cases therefore does not guarantee accurate interpolation between them. This difficulty is particularly pronounced for nonequilibrium targets such as stress, heat flux, and velocity-distribution functions, whose errors are more sensitive to the flow regime than those of conservative macroscopic variables29.
One approach to handling high-dimensional flow data is to represent each solution using a small number of latent variables. Closely related linear methods such as principal component analysis (PCA)30 and proper orthogonal decomposition (POD)31,32 construct these variables from dominant global modes and provide a clear hierarchy for truncating the representation. Their efficiency may nevertheless decrease when the flow solutions vary along a strongly nonlinear manifold. An early connection between modal reduction and machine learning was established by Milano and Koumoutsakos, who interpreted POD as a particular linear neural-network representation and extended it using a nonlinear neural network for the reconstruction and prediction of near-wall turbulent flows33. This viewpoint motivates modern autoencoders and variational autoencoders (VAEs), which learn nonlinear mappings between physical fields and low-dimensional latent variables34. However, convolutional implementations generally assume a fixed grid and a fixed arrangement of output channels. Neural fields take a different approach by representing a field as a continuous function of its coordinates35. Fourier features help these models capture small-scale spatial variations36, while auto-decoders assign an optimized latent code to each observed field without requiring a separate encoder[37]. This coordinate-based formulation can be applied to fields stored on different spatial discretizations[38, 39] [19]. However, latent codes learned jointly with a decoder are not unique: they can be rotated or rescaled without changing the reconstructed fields, and distances between latent codes do not necessarily measure differences between the decoded flow solutions[40]. Consequently, accurate reconstruction alone does not guarantee that the learned latent space is well organized for predicting solutions at unseen physical conditions.
A second line of research concerns the prediction of solution representations at previously unseen physical conditions. Deterministic operator-learning and conditional-field models approximate a single-valued map from governing inputs to solution fields17–19. Conditional generative models instead provide a sample-based formulation by transforming samples from a tractable reference distribution into solution states under a specified physical condition. Unlike a deterministic regressor that returns a single point estimate for each condition, this construction retains a sampling dimension and can represent conditional variability or multiple plausible solution states when such variability is resolved by the training data. This capability should not, however, be interpreted as physical uncertainty when only one converged reference solution is available at each condition; in that setting, the generated spread primarily reflects the learned transport and sampling procedure. In diffusion models, data are progressively perturbed according to a prescribed stochastic process, and generation is performed by numerically solving learned reverse-time dynamics conditioned on the physical parameters[41, 42, 43]. Their performance can be sensitive to the noise schedule, prediction parameterization, conditioning mechanism, and numerical sampler. Flow matching (FM) instead learns a time-dependent conditional vector field along a prescribed probability path and generates samples by integrating an ordinary differential equation (ODE)[44, 45, 46]. It replaces reverse-time denoising with ODE transport and may reduce the number of model evaluations, although its accuracy remains sensitive to the source–target coupling, path geometry, and integration error. Conditional diffusion and flow-based models have recently been investigated for parametric equations, physical forecasting, field reconstruction, and continuum-flow prediction[47, 48, 49, 50, 51]. Existing comparisons are nevertheless often confounded by differences in field representation, conditioning architecture, training data, and evaluation protocol. Moreover, for deterministic parametric problems with only one reference state at each sampled condition, it remains unclear whether conditional generative transport improves prediction beyond a deterministic conditional regressor or primarily provides an alternative mechanism for navigating the learned solution manifold.
A third line of research seeks to incorporate physical information into learned solution models. Physics-informed neural networks introduce governing equations, initial or boundary conditions, and other physical relations into the training objective through differentiable residual terms[52]. Weak formulations can improve robustness when strong-form derivatives are unreliable or when the target fields contain sharp spatial variations[53]. For diffusion models, physical information has also been introduced through additional training losses, conditional guidance, or corrections applied during sampling[54, 55, 56]. Such methods can reduce selected equation or boundary residuals, but a lower residual does not by itself establish that every generated field satisfies the complete governing system. The result depends on whether all variables required by a constraint are available, whether the corresponding derivatives can be evaluated accurately, and whether residuals with different scales are normalized appropriately. Moreover, increasing the weight of a physical penalty can improve the targeted diagnostic while moving predictions away from the data-supported solution manifold, whereas a weak penalty may have little measurable effect. Parameter-efficient adaptation provides a general mechanism for limiting such changes: a pretrained backbone is frozen and only a comparatively small residual module is optimized for the downstream objective[57, 58]. These techniques were developed for controlled model adaptation rather than for enforcing physical laws, and their use alone therefore provides no guarantee of conservation or kinetic consistency. The unresolved problem is how to impose only those physical relations that can be evaluated from the available outputs, restrict the capacity of the resulting correction, and improve the targeted physical diagnostics without materially degrading the predictive accuracy of the pretrained generative surrogate.
Taken together, these studies leave a coupled representation–generation–consistency problem. Existing approaches commonly address field compression, conditional prediction, and physical regularization as separate tasks, so an improvement at one stage does not necessarily persist through the complete parameter-to-field pipeline. In particular, it remains unclear whether a coordinate-based decoder can provide a common compact representation for heterogeneous rarefied-flow outputs, whether diffusion and flow matching can accurately transport that representation across sparsely sampled physical conditions under a common evaluation protocol, and whether selected physical diagnostics can subsequently be improved without retraining or materially perturbing the data-supported generative backbone. Resolving these questions requires reconstruction, conditional-generation, and physics-correction errors to be evaluated separately, because the final field error may otherwise conceal the stage at which predictive accuracy is lost.
To address these questions, we develop a conditional latent generative framework and evaluate it on two steady two-dimensional rarefied-flow benchmarks: DVM-resolved lid-driven cavity flow parameterized by Kn and DSMC-resolved cylinder flow parameterized by (Kn, Ma)29. The central contribution is a unified neural-field representation–conditional latent generation–frozen physics adaptation framework, together with an evaluation that separates the evidence for each component. First, coordinate-based neural-field auto-decoders are trained without supplying the physical conditions to the representation stage, and principal-component charts fitted exclusively to the training lookup codes provide compact coordinates for subsequent conditional transport. Second, conditional flow matching and diffusion act directly on these latent charts and generate condition-dependent terminal latents from reference or source latent samples. Their frozen pipelines are compared across a one-dimensional condition coordinate and a sparse two-dimensional parameter geometry, using common data partitions and decoded-field metrics within each benchmark. A direct condition-to-chart MLP is additionally decoded by the same field representation to test whether the sampling-based formulation improves steady point prediction. Third, generator-specific structured low-rank adapters introduce capacity-limited physical corrections while leaving both the selected generator and the neural-field decoder frozen. For the cavity flow, parallel adapter branches modify the four macroscopic outputs , the velocity-basis coefficients, and the corresponding nonequilibrium bias terms, while preserving the decoder’s positive and bounded distribution parameterization. For the cylinder flow, the adapter corrects the standardized density, streamwise-velocity, transverse-velocity, and temperature outputs and includes a wall-localized normal-velocity branch together with a no-penetration wall map. All learned adapter output projections are initialized to zero, but the FM and diffusion branches follow the same freezing principle rather than an identical optimization protocol: their ranks, latent exposure, sampling density, and checkpoint criteria are selected separately. Finally, the evaluation distinguishes auto-decoder reconstruction error, conditional-generation error, physics-correction error, interpolation, and out-of-support extrapolation. The resulting evidence supports accurate interpolation over covered parameter ranges and shows that frozen structured adaptation can improve explicitly computable physical diagnostics with limited changes in field accuracy. The evaluation is designed to answer, rather than assume, the question of where sampling-based transport adds value: it includes a deterministic condition-to-chart MLP baseline under the same frozen decoder, five-seed repetitions of every generator, and frozen physics-adaptation audits, so that the regime of applicability of each formulation is established under common data partitions and decoded-field metrics.
This paper is organized as follows. Section II defines the two rarefied-flow problems, describes the DVM and DSMC datasets, and establishes the common prediction and evaluation protocol. Section III presents the neural-field auto-decoders, train-only latent charts, conditional FM and diffusion models, and structured physics adaptation. Section IV reports the controlled representation, generator, sampler, and physics-adapter ablations and fixes the models used thereafter. Section V evaluates representative interpolation and extrapolation cases for the cavity, assesses additional cylinder conditions, benchmarks the frozen generators against direct latent MLPs, and discusses the cross-benchmark behavior. Finally, Section VI summarizes the principal findings, limitations, and directions for further work.
Problem setup and data generation
Lid-driven cavity flow and DVM data
Physical configuration and kinetic model
We consider the steady flow of a monatomic gas in the square cavity , where is the cavity side length. The dataset is expressed in nondimensional units with , reference density , wall temperature , and gas constant , so that the reference most probable speed is . The upper wall moves in the positive direction at , while the remaining walls are stationary. All four walls are isothermal and use fully diffuse Maxwell reflection; the density of the incoming wall Maxwellian is chosen to give zero net normal mass flux. The flow is two-dimensional in physical space: there is no dependence on a third spatial coordinate and the third macroscopic velocity component is zero. Molecular translation nevertheless remains three-dimensional for the monatomic gas, which therefore has specific-heat ratio . The cavity Knudsen number is , where is the reference mean free path.
Let denote the particle velocity-distribution function, where are the time and physical coordinates and are the particle-velocity components associated with transport in the two-dimensional physical plane. The remaining component is retained in because the particle velocity is three-dimensional, but it produces no third-direction spatial transport and is not explicitly discretized. The reference solutions are governed by the BGK model[59],
where is the relaxation time and the local three-dimensional Maxwellian is
The quantities , , , and in Eq. (2) are, respectively, the local mass density, the two in-plane macroscopic velocity components, and the translational temperature obtained from moments of . The saved fields are steady solutions of Eq. (1). For the parameters used by the solver, the viscosity model and relaxation time reduce to
where is the local dynamic viscosity, is its reference value, is the local pressure, and is the viscosity exponent. The value is inherited from the benchmark DVM configuration29, rather than introduced here as a new gas-model calibration. Consequently, is the only physical condition varied across the cavity cases.
Reduced distributions and macroscopic moments
Because the physical problem has only two spatial directions, the DVM explicitly discretizes only the in-plane particle velocities . The contribution of the remaining particle-velocity component is retained by integrating it into two reduced distributions, for which we use the paper notation
The corresponding equilibrium distributions are
and integration of Eq. (1) over the hidden velocity gives
The conserved moments are recovered by integration over the explicit two-dimensional velocity plane:
where is the total energy. With the in-plane peculiar velocity components , the pressure tensor and heat flux are
and
where and denote the in-plane stress tensor and heat flux, respectively. The DVM dataset provides the conservative-state vector , , , and the two reduced distributions. The learned field representation predicts the four primitive variables together with and ; conservative and higher-order quantities can then be recomputed using the same velocity quadrature.
DVM discretization and data partitions
The reference fields were obtained with a finite-volume discrete velocity method3,29. The physical domain uses a uniform cell-centered Cartesian grid, whose coordinates range from to in each direction. The nominal velocity domain is and is represented by an uniform cell-centered grid with nodes from to . Its spacing is , and velocity moments are evaluated by the composite midpoint Newton–Cotes rule,
where is an arbitrary velocity-space integrand, and and index the discrete nodes in the and directions, respectively. Thus, one reduced distribution contains values per condition, and the pair contains values. This dimensionality motivates a continuous coordinate representation rather than direct generation of the full physical–velocity tensor.
The representation and conditional generators use the following 16 training conditions:
The three conditions are used for architecture, hyperparameter, and checkpoint selection and are therefore validation rather than blind-test cases. For concise presentation, the main-text field comparisons use the following four conditions from the additional evaluation data:
where are interpolation cases inside the training support, whereas and probe lower- and upper-side extrapolation, respectively. These four conditions are selected as representative predictions for the main-text comparison.
Rarefied flow around a cylinder and DSMC data
Physical configuration and operating conditions
The second benchmark considers a fixed circular cylinder in a high-speed rarefied argon stream. This canonical configuration contains a detached bow shock, a near-wall Knudsen layer, and a downstream wake whose structures vary strongly with compressibility and rarefaction29. The cylinder is centered at the origin and has diameter and radius . Coordinates are nondimensionalized by , giving the computational box
with a periodic spanwise direction in the underlying DSMC calculation. The wall and freestream temperatures are both , and gas–surface interaction at the cylinder is modeled by fully diffuse reflection with complete thermal accommodation.
The operating condition is described by
where Kn and Ma are the Knudsen and Mach numbers; , , and are the freestream mean free path, speed, and temperature, respectively; is the freestream sound speed, is the Boltzmann constant, and is the argon atomic mass. The source simulation campaign spans in increments of and in increments of , yielding 2400 operating conditions29. The present study uses the following tensor-product subset of 64 cases:
This range spans substantial changes in the bow shock, near-wall Knudsen layer, and downstream wake: decreasing Kn produces thinner and sharper structures, while increasing rarefaction makes the macroscopic features more diffuse.
DSMC generation and macroscopic fields
The reference fields were generated using the direct simulation Monte Carlo method1,29, with a variable-hard-sphere model for argon. Hierarchically refined Cartesian cells resolve the cylinder surface, near-wall rarefaction, bow shock, and wake. After a statistically steady state is reached, particle statistics are time averaged to reduce DSMC sampling noise. The DSMC data were generated using the open-source SPARTA solver with the variable-hard-sphere (VHS) collision model for argon and the no-time-counter collision scheme, following the simulation protocol of TransportBench29; the argon collision and reference parameters are summarized in Table I.
| Quantity | Value |
| Atomic mass | |
| VHS reference diameter | |
| VHS viscosity index | |
| Reference temperature | |
| Wall and freestream temperature |
Table I. Argon collision and reference parameters of the DSMC benchmark1,29.
The nominally two-dimensional prediction task retains number density , the in-plane velocity components , and translational temperature . Rotational and vibrational temperatures are absent for monatomic argon, and the spanwise mean velocity is not included as a prediction target. Mass density and pressure are not treated as independent learned variables but are reconstructed through
The native multilevel Cartesian mesh is retained rather than interpolating the training fields to a common image grid. Its resolution varies with Kn and is locally refined around the cylinder, bow shock, and wake. Only fluid cells contribute to field losses and metrics. Regular grids used later for common-resolution field comparisons are coordinate queries to the learned neural field, not the representation on which the decoder is trained. Concretely, all cylinder out-of-sample field comparisons are evaluated on a regular common grid of uniform points spanning , the cavity error metrics are evaluated on the native DVM grid with its velocity quadrature (Eq. (12)), and the displayed cavity contour figures use uniform grids.
Preprocessing and data partitions
For each condition, the four learned targets are expressed as disturbances relative to the prescribed freestream state,
where the subscript denotes the prescribed freestream value and the superscript denotes vector transpose. Each component is subsequently divided by a case-balanced root-mean-square (RMS) scale fitted from training cases only. This construction makes zero disturbance exactly equal to the prescribed freestream state.
The 64-case tensor grid is divided into 48 training, eight validation, and eight internal test conditions. The validation set is
and the internal test set is
where each ordered pair denotes . Neither validation nor test cases contribute to target normalization or neural-field parameter optimization. For concise presentation of the out-of-sample predictions, the main text displays the following representative conditions:
where each ordered pair denotes .
Unified prediction task and evaluation protocol
Conditional prediction and error decomposition
Both benchmarks are formulated as conditional parameter-to-field prediction. Let denote the physical condition and the coordinates at which a field is queried:
The cavity target comprises the primitive variables and reduced distributions , whereas the cylinder target is ; the subscripts and identify the cavity and cylinder benchmarks, respectively. Additional macroscopic quantities are obtained from Eq. (18). Throughout the remainder of the paper, boldface denotes a multichannel target or prediction, scalar remains the second Cartesian coordinate, and is reserved exclusively for the heat-flux component defined in Eq. (11).
For either benchmark, let denote the conditional generator and the frozen neural-field decoder, with and denoting their parameter sets. The generator maps the th reference or source sample to a compact latent-chart coordinate,
The data provide one converged steady reference field at each condition. Accordingly, the primary point prediction is obtained by averaging the chart endpoints before the fixed inverse transformation and nonlinear decoding,
where is the inverse latent-chart transformation, a hat denotes a generated endpoint, and is the ensemble size. This convention is fixed for all reported methods; because the decoder is nonlinear, it is not equivalent to averaging independently decoded fields. The endpoint spread is treated as a diagnostic of the learned transport, not as calibrated physical uncertainty.
A frozen-decoder audit quantifies this noncommutativity (Section A): the relative difference between latent-space and decoded-space averaging is at most for FM and, away from the lower extrapolation condition, at most for diffusion; at latent averaging contributes a non-negligible part of the extrapolation error.
Three sources of error are reported separately. The representation error is obtained by decoding the optimized lookup latent of an observed case; it measures the neural field’s attainable reconstruction floor. The conditional-generation error is obtained from Eq. (24) with the decoder fixed and therefore includes latent-transport error. The physics-correction effect is the paired change from a frozen generative prediction to its adapter-corrected counterpart. Because the decoder and correction are nonlinear, these quantities are not assumed to form an additive error decomposition.
Field-error measures
For the cavity, let index physical grid points, index discrete velocity nodes, index the four conservative-state components, and denote the DVM quadrature weight. The macroscopic and kinetic relative errors are
with and a small preventing division by zero. For a derived tensor or vector , the componentwise relative error is
where indexes the pressure-tensor or heat-flux components defined in Eqs. (10) and (11). The matched-grid reduced-BGK diagnostic is evaluated with the same physical grid, velocity quadrature, and discrete differentiation rule for every compared model. It is therefore used for paired comparisons and is not interpreted as a rigorous bound on the continuous BGK residual.
For the cylinder, each reported field is first nondimensionalized by its physical reference scale. Let denote a generic dimensionless field from the following set:
For a condition indexed by , the area-weighted root-mean-square error (RMSE) is
where indexes spatial samples and is the corresponding spatial quadrature weight, which is uniform on the common regular grid used for the final field comparison.
Data use and reporting conventions
Training conditions determine the field decoder, latent chart, conditional generator, and train-only normalization statistics. Validation conditions are used for architecture, hyperparameter, and checkpoint selection. The final out-of-sample conditions do not contribute to gradient updates or checkpoint ranking. A condition is classified as interpolation when it lies within the support of the training conditions and as extrapolation when it lies outside that support; the two categories are reported separately.
Errors are first computed for each physical condition and are then aggregated with equal case weight,
where is the error for the th condition and is the number of conditions included in the reported aggregate. This prevents conditions with finer DSMC meshes or more spatial samples from dominating the summary. Frozen FM and Frozen Diffusion denote the unadapted conditional generators. A physics-constrained label is reserved for an adapted checkpoint that passes all benchmark-specific feasibility gates. The cavity FM checkpoint satisfies its feasibility gates; the corresponding diffusion checkpoint and both retained cylinder checkpoints pass their respective gates. All models are compared against the same DVM or DSMC reference within each benchmark. Physics residuals are reported relative to the corresponding frozen generator and are not compared as absolute values between the cavity and cylinder problems, because the available variables and evaluable physical constraints differ.
Conditional latent generative framework
Overall parameter-to-field architecture
The proposed framework separates high-dimensional field representation from conditional generation. For each benchmark, a neural-field auto-decoder first assigns an independently optimized lookup code to every training condition and learns a coordinate decoder . The physical condition is not supplied to this representation stage; hence, the case-dependent information required for reconstruction must pass through . A train-only latent transformation then maps the lookup codes to a lower-dimensional chart on which the conditional generative models are trained. This separation avoids requiring either flow matching or diffusion to generate the full spatial or physical–velocity tensor directly.
Let index the generator family and denote its model-specific reference state. A conditional generator produces a latent-chart sample, which is mapped back to the auto-decoder space and evaluated at arbitrary query coordinates:
where and denote the chart and auto-decoder coordinates, respectively. For diffusion and the cavity FM model, is sampled from a tractable Gaussian reference distribution. For the cylinder FM model, it instead contains an observed source latent and its source condition; the model transports this anchor to the requested target condition. These two choices instantiate the same reference-to-target formulation but are not assumed to define identical probability paths. Their training and inference are specified separately below.
During conditional-generator training and evaluation, and are frozen. Consequently, differences between FM and diffusion cannot be absorbed by retraining the field representation. The latent-endpoint ensemble mean is used to obtain the primary steady-field prediction, while the sample spread is retained only as a transport-sensitivity diagnostic, as specified in Section II C. The sampling dimension is therefore preserved without interpreting it as a measured physical uncertainty distribution.
The optional physics-correction stage also preserves the selected generator and decoder. It augments designated decoder heads with a structured low-rank adapter ,
where denotes the only learned correction parameters. The upward adapter projections are initialized to zero, so the learned low-rank branches initially make no change to the frozen prediction. The cylinder protocol may additionally apply a prescribed hard wall-normal projection; the exact identity at initialization refers to the learned adapter before this deterministic map. Separate adapters are trained for FM and diffusion. They share the same freezing principle and benchmark-specific output structure but not necessarily the same optimization schedule or selected checkpoint.
For the cavity problem, the auto-decoder has a lookup code of dimension 24 and a train-only PCA chart of dimension four. Its spatial heads predict , while coupled spatial–velocity heads predict the coefficients and biases used to construct the positive reduced distributions and . The physics adapter acts in parallel on the four primitive outputs, the velocity-basis coefficients, and the corresponding nonequilibrium biases. Thus, both macroscopic and kinetic predictions remain linked through one generated latent state.
For the cylinder problem, the auto-decoder uses a 32-dimensional lookup code and a 12-dimensional latent chart. Its coordinate network predicts the four standardized freestream disturbances in Eq. (19); density and pressure are recovered through the hard algebraic relations in Eq. (18). The structured adapter corrects these four macroscopic decoder channels and includes a wall-localized normal-velocity branch. The retained cylinder FM generator is a source-to-target condition transport over local observed anchors and uses a decoder-metric PCA chart, whereas its diffusion counterpart generates an ordinary PCA chart of the same retained dimension from a Gaussian reference. Both are mapped back to the same 32-dimensional lookup-latent space and decoded by the same frozen neural-field representation.
No deterministic condition-to-latent regressor is used inside either generative pipeline. Such a regressor is retained only as a controlled baseline and does not initialize, supervise, or correct the generated latent samples. In particular, the term “plain-concat MLP” used for one FM architecture, where MLP denotes multilayer perceptron, refers to a parameterization of the time-dependent velocity field; it is not a direct single-valued mapping from the physical condition to the terminal latent. For the controlled deterministic baseline, an ensemble of regressors maps the normalized physical condition directly to the retained latent chart,
where is the number of independently initialized MLP members. The cavity baseline has , width 128, three hidden layers, and Knudsen Fourier frequencies 0.5 and 1.0; it predicts the common whitened PCA-4 chart. The cylinder baseline has , width 64, and two residual blocks; it predicts the ordinary unwhitened PCA-12 chart. Both models, their train-only chart transformations, and the field decoders are frozen before evaluation on the additional condition. They are comparison models only and are not components of either generative pipeline. The same protocol is additionally executed with five independent seeds per configuration, at the parameter-matched capacities, and, for the cylinder, on both retained charts; these controlled extensions are reported in Section V A. The complete separation between representation learning, latent generation, and coordinate-based field reconstruction is summarized in Figure 1.
Figure 1. Conditional latent generative framework. Conditions are excluded from representation training. FM and diffusion operate in train-only latent charts using Gaussian references, except for the observed-anchor cylinder FM. Mean endpoints are inverse transformed and decoded before a structured adapter corrects selected channels with the upstream pipeline frozen.
Neural-field auto-decoder and latent representation
Coordinate-based field decoders
For both benchmarks, the representation model is an auto-decoder rather than an encoder–decoder[37]. A trainable lookup vector identifies the th observed condition, while a shared coordinate network represents the associated continuous field,
Neither nor is provided to . The condition labels are therefore unavailable as a shortcut for field reconstruction, and all case dependence must pass through . The two decoders share this principle but use different coordinate features and output factorizations because the cavity includes velocity-distribution functions whereas the cylinder data contain macroscopic fields on nonuniform meshes.
Let denote a Fourier encoding of a normalized coordinate vector ,
The encoded coordinates are processed by residual multilayer perceptrons. In a latent-conditioned block, feature-wise linear modulation (FiLM)[60] is applied after layer normalization,
where indexes the residual blocks, denotes layer normalization, denotes componentwise multiplication, and are learned feature-wise scale and shift maps, respectively, is a two-layer residual mapping with the sigmoid linear unit (SiLU) activation, and is a residual scaling factor. The FiLM projections are initialized so that at initialization.
For the cavity, the coordinates are normalized as , , , and . The principal spatial branch uses frequencies , supplemented by smooth indicators of the moving lid, , and of the stationary walls, . A second spatial encoding with frequencies supplies higher-frequency features to the kinetic branch. The velocity encoder takes the feature vector together with Fourier frequencies . The resulting feature dimensions are 32, 18, and 28 for the principal spatial, kinetic spatial, and velocity encoders, respectively. A residual scale of is used throughout the cavity decoder. The lookup code modulates the even-numbered blocks of the principal spatial trunk and every block of the kinetic spatial branch.
A width-640 spatial trunk with 12 residual FiLM blocks predicts the standardized primitive outputs (). A width-768, eight-block kinetic spatial branch produces coefficient vectors and scalar biases . Independently, a condition-invariant width-512 velocity trunk with 10 residual blocks produces the two velocity bases . The raw nonequilibrium corrections are
and are smoothly bounded before being applied to the local equilibrium:
This construction preserves , couples the kinetic outputs to the predicted macroscopic state through and , and avoids an unconstrained exponential correction. The cavity decoder has a 24-dimensional lookup code and approximately 26.37 million trainable parameters.
For the cylinder, the coordinate encoder combines normalized Cartesian coordinates with cylinder-centered polar geometry. Its 163 features include geometric localization terms together with the Cartesian, polar-angle, and wall-distance frequency sets and the wall-localization length scales (in units of ),
The remaining geometric inputs include the signed distance , the polar angle, the wall-normal direction, and exponential wall-distance localization features. These features distinguish the near-wall Knudsen layer from the detached shock and wake without first interpolating the native DSMC cells to an image grid.
The cylinder decoder uses a 32-dimensional lookup code, a width-512 trunk, and 12 FiLM residual blocks, with the lookup code injected through a scaled input projection and through FiLM in every block. Four independent width-192 heads predict the standardized components of in Eq. (19); separate heads reduce interference between density, the two velocity components, and temperature. Their final layers are initialized to zero, making the initial prediction the prescribed freestream state. The physical variables are reconstructed by inverting Eq. (19), after which and are imposed algebraically. This decoder has approximately 7.22 million trainable parameters. The principal architectural settings of both neural fields are summarized in Table II.
| Component | Cavity DVM | Cylinder DSMC |
| Lookup-latent dimension | 24 | 32 |
| Principal trunk | 640 × 12 blocks | 512 × 12 blocks |
| Additional branch | 768 × 8 kinetic | Four 192-wide heads |
| Velocity trunk/basis | 512 × 10, rank 256 | Not applicable |
| Coordinate feature dimensions | 32/18/28 | 163 |
| Trainable parameters | 26.37 M | 7.22 M |
Table II. Neural-field auto-decoder architectures used for the two benchmarks. The listed parameters include the lookup table during representation training.
Lookup-latent optimization
The decoder parameters and the lookup table are optimized jointly over the training cases:
The physical condition is excluded from Eq. (42); it is used only after representation training to learn conditional transport. Small latent-norm and centered-table penalties remove unconstrained drift without assigning a physical meaning to individual raw latent coordinates.
For the cavity, reconstruction combines normalized conservative and primitive losses, the quadrature-weighted relative losses of and , and consistency between the conservative state predicted by the macroscopic head and the moments of the kinetic heads:
where and are normalized smooth- losses; and have the quadrature-weighted relative- form of Eq. (27); and compares the four discrete moments of with the macroscopic head. The terms and regularize the lookup table. A normalized second-difference penalty is applied along , but no first-difference penalty is used. The scale-invariant basis-correlation penalty is ramped from zero between 50% and 75% of training to a maximum coefficient of . provides a weak diffuse-wall preconditioner for the incoming portions of and .
Cavity representation training contains three phases within one 200,000-update optimization. Updates 1–2,000 use and the lookup-code regularizers. Updates 2,001–60,000 introduce the full-grid kinetic losses, change the macroscopic weights to 0.75 and 0.25, and periodically evaluate . The remaining updates additionally activate the latent-curvature, basis-correlation, and wall-preconditioning terms in Eq. (43). Each kinetic update uses all 16 training cases, all 2,500 spatial cells, and all 6,400 velocity nodes. Consequently, the principal losses are full-grid quantities rather than Monte Carlo estimates; only the moment and wall terms are evaluated at their prescribed intervals to control memory and computation.
For the cylinder, the primary data loss is an area- and sampling-confidence-weighted smooth- loss on the standardized freestream disturbances,
where , , and index cases, sampled cells, and the four output channels, respectively; the SmoothL1 term is the Huber (smooth ) loss with threshold 0.02; is the corresponding number of averaged elements; contains train-only, case-balanced root-mean-square (RMS) scales; and is the normalized product of cell area and a clipped square-root particle-confidence factor. Each update samples 16 cases and 18,432 points per case, with 20% global, 25% shock, 30% near-wall, 20% wake, and 5% far-field samples. This sampling distribution emphasizes the shock and Knudsen layer during optimization, whereas validation remains area weighted over the complete native mesh.
The cylinder representation objective before the late physical refinement is
where follows a half-cosine ramp from zero after 25% of training to at the final update. The local affine-surface term uses neighboring points in only to regularize the geometry of the lookup table; the labels still do not enter the decoder. A late, weak representation-stage physical refinement applies normalized wall no-penetration, inlet freestream, and control-volume mass terms in the ratio 0.30:0.15:0.55, with a combined peak coefficient of . Its multiplier is zero through 68% of training, reaches one at 82%, and, after 92%, decays smoothly to 0.10 at the final update. This preconditioning is distinct from the generator-specific physics adapters introduced later: it forms part of the common frozen field representation and does not use generated latents. The cylinder model is trained for 180,000 updates using region-aware point sampling.
For both benchmarks, a held-out Oracle code is obtained by freezing and optimizing only a new against the observed field. Oracle decoding measures the representational floor of the frozen neural field; it is not a deployable inference procedure and is kept separate from conditional-generation error.
Train-only latent charts
Raw auto-decoder codes possess a gauge freedom: an invertible change of latent coordinates can be compensated by the decoder without changing the represented fields. The generative models therefore operate on a train-only principal component analysis (PCA) chart30 rather than directly on individual lookup coordinates. This choice removes weakly constrained directions, restricts generated endpoints to the empirically supported latent subspace, and balances coordinate scales for FM and diffusion optimization. It thereby reduces the statistical and numerical burden of learning all 24 or 32 raw lookup coordinates from sparsely sampled physical conditions. Directly transporting the raw lookup coordinates would force the generator to fit directions that the decoder can absorb without changing the represented field, which is ill-conditioned under 16 or 48 training cases; the chart removes these directions before transport. The chart is not a separate field representation: it is a low-dimensional coordinate system for conditional generation, and every generated endpoint is mapped back to the original auto-decoder coordinates before evaluation by the frozen decoder. Let be the number of training conditions, the lookup dimension, , and the matrix whose th row is . Its singular value decomposition is
where and contain the left and right singular vectors, contains the singular values, contains the first right singular vectors, and is the retained, unscaled PCA score. Every centering, scaling, metric, and basis quantity is fitted using training codes only and is refitted when a training condition is omitted during model selection.
For the cavity, . Its PCA scores are whitened using
where is the empirical standard deviation of the th PCA score. The whitening floor prevents weak lookup-code directions from being amplified. Candidate ranks are assessed by decoding the truncated lookup codes, so rank selection depends on field fidelity in addition to explained latent variance.
The cylinder diffusion branch uses the rank-12 unwhitened PCA scores in Eq. (46). Before generative optimization, each retained score is standardized by its train-only standard deviation with a lower scale of . This scaling is an optimizer preconditioner and is undone before the ordinary PCA inverse maps the generated endpoint to the 32-dimensional lookup-latent space.
The retained cylinder FM branch uses a separate rank-12 decoder-metric PCA chart. Frozen-decoder sensitivities are evaluated at global, shock, near-wall, wake, and far-field points to estimate one Gauss–Newton matrix per training case. Their mean is symmetrized and divided by its mean eigenvalue (equivalently, its trace divided by ) to obtain in the 32-dimensional lookup space. It is blended with the Euclidean metric and factorized by its symmetric positive-definite square root,
PCA is then applied to , followed by the same score-scale floor of . Euclidean displacement in this chart consequently approximates, locally, a standardized decoded-field change and gives greater weight to lookup directions to which the frozen decoder is sensitive. The metric and PCA basis are recomputed from retained cases within every strict leave-one-condition-out fold.
Thus, the cylinder FM and diffusion models share the same frozen decoder, 32-dimensional lookup space, retained chart dimension, data partitions, and decoded-field evaluation, but they do not use an identical chart basis. The distinction is retained because decoder-metric coordinates belong to the retained curved-transport FM architecture, whereas the retained diffusion baseline uses the ordinary PCA chart. In all cases, is a fixed linear transformation during generator and physics-adapter training.
Conditional flow matching
Flow matching (FM) learns a conditional velocity field that transports a reference state to a target latent without requiring likelihood evaluation[44, 46]. Let denote a target chart coordinate, a reference or source coordinate, and the transport time. For a prescribed conditional path
the conditional vector field is trained by
where denotes the FM-network parameters, overdots denote derivatives with respect to , and is the chart dimension. Generation integrates the neural ordinary differential equation
The cavity and cylinder models use different couplings because the former has a one-dimensional condition space with a Gaussian reference, whereas the latter transports observed solution anchors over a two-dimensional condition manifold.
Cavity Gaussian-reference flow matching
For the cavity, and is the four-dimensional whitened PCA coordinate of an observed lookup latent. The selected probability path is
The residual Gaussian scale at avoids numerically forcing a continuous reference law to a strictly singular endpoint.
The condition is normalized using training-set endpoints,
so on the training interval. Fixed Fourier encodings use for time and for the condition. The vector-field input
is processed by a plain-concatenation MLP comprising an input layer of width 704, six width-704 SiLU layers, layer normalization, and a four-dimensional zero-initialized output. It contains 2,996,932 parameters. A weak second-difference penalty in the normalized Knudsen coordinate is
where and . This term is evaluated in 32-bit floating-point arithmetic (FP32), giving
The principal FM regression is evaluated with 16-bit brain floating-point (BF16) mixed precision. Inference uses a 48-step explicit Heun solver and antithetic Gaussian initial states. For even ensemble size , antithetic sampling draws independent states and appends their negatives, so the empirical reference mean is exactly zero before transport. The ensemble-mean endpoint is inverse transformed and decoded; 64 endpoints are used for validation and 256 for the final field comparison.
Cylinder curved condition-manifold transport
For the cylinder, the source is an observed chart coordinate at , and the target is at . Each condition is standardized componentwise as , using training-set means and standard deviations. Training uses bidirectional edges between four nearest conditions and the rectified path
To represent curvature in the two-dimensional condition manifold without an unrestricted source–target lookup, the velocity is decomposed as
where
Here supplies the complete first-order tangent response, is a state-, time-, and condition-dependent scalar coefficient, and the eight learned vectors span a low-rank quadratic correction. Both output heads are zero initialized, and identically for . The shared residual MLP has width 128, three residual blocks, expansion factor two, SiLU activation, layer normalization, and time frequencies ; no condition Fourier features are used. It contains 208,144 trainable parameters.
Let denote the endpoint obtained by numerically integrating Eq. (51) from source to target condition. The endpoint and consensus terms used for cylinder training are
where , is the number of sampled source anchors for target , and is their endpoint mean. For a mini-batch of target cases, the decoded-field term is
where the inverse chart used to obtain and the decoder are fixed; and are the numbers of cases and sampled spatial points in this auxiliary batch, and superscript std denotes the standardized cylinder outputs. The complete objective is
The auxiliary terms are estimated every 10 updates after the first 25% of optimization, using a differentiable four-step Heun integration. The field term samples global, shock, near-wall, wake, and far-field regions with respective fractions 30%, 25%, 25%, 15%, and 5%, and includes a worst-case contribution in addition to the mean. Gradients pass through the frozen decoder to the generator, but no decoder parameter is updated.
At inference, a local nondegenerate triangle containing the query is selected in condition space. Its three vertices independently generate target latents with a 24-step Heun solver, after which their nonnegative barycentric weighted mean is used. Outside the convex hull, three nearest anchors and positive distance-softmax weights are used. Small centered source perturbations of scale 0.01 provide a transport-sensitivity ensemble; they do not represent measured DSMC uncertainty.
Conditional diffusion
The conditional diffusion models act on standardized chart coordinates[41, 42, 43]. A continuous linear variance-preserving (VP) schedule is defined by
where is a clean target chart coordinate and . Using to distinguish a diffusion regression target from the physical field , the three parameterizations are
with
where denotes the denoiser parameters. The subscript follows the conventional name “x-prediction” and means prediction of the clean latent ; it is unrelated to the spatial coordinate . Although these parameterizations describe the same forward perturbation analytically, their loss weighting across noise levels and their numerical conditioning differ.
Both denoisers concatenate the noisy chart state, a Fourier time encoding, and the normalized physical condition. The cavity model additionally applies Fourier features at condition frequencies ; the cylinder model uses standardized directly. The cavity denoiser has the same width-704, six-layer plain-concat architecture and parameter count as its FM counterpart. Its retained formulation uses -prediction, FP32 training, and . The cylinder denoiser is a 203,404-parameter residual MLP with width 128, three blocks, expansion factor two, and time frequencies . It uses -prediction and .
Because each cylinder condition has one reference solution, the cylinder fit also constrains the endpoint mean and suppresses artificial conditional spread. For antithetic samples at each of four sampled conditions,
These endpoint terms are estimated with four-step denoising diffusion implicit model (DDIM) sampling[61] every 20 updates after 60% of optimization. The cavity diffusion fit does not use endpoint regularization. During cylinder sampling, the predicted clean chart coordinate is smoothly clipped at magnitude 6 before each reverse update.
The ablation protocol includes the denoising diffusion probabilistic model (DDPM) reverse chain[42], deterministic DDIM sampling, and the second-order multistep (2M) DPM-Solver++ method[62]. The frozen cavity model is sampled with 35 DPM-Solver++(2M) steps, using 64 antithetic samples for validation and 256 for final evaluation. The frozen cylinder model uses 40 deterministic DDIM steps and 40 antithetic samples. The clean chart endpoint is averaged before the fixed inverse transform and nonlinear field decoding. Generated spread is reported as a sampler and transport diagnostic, not as calibrated physical uncertainty.
Controlled FM–diffusion comparison
Within each benchmark, FM and diffusion use the same representation-stage lookup table, field decoder, training/validation partition, target normalization, and decoded-field metrics. For the cavity, the comparison is additionally parameter matched: both conditional networks contain 2,996,932 parameters, operate on the same rank-four whitened PCA chart, use the same condition features, and are evaluated with antithetic ensembles. The remaining differences are the training objective, probability path or noise schedule, numerical precision selected by ablation, and deployment integrator.
For the cylinder, the two conditional networks have similar capacity and the same retained chart dimension, but the comparison is not representation identical. Diffusion uses the ordinary rank-12 PCA chart, whereas the retained FM uses the decoder-metric rank-12 chart in Eq. (48). FM also transports local observed anchors and uses barycentric aggregation, while diffusion starts from a Gaussian reference. These differences are retained because they belong to the frozen retained models; they are disclosed rather than absorbed into a claim of perfectly matched generative computation. All endpoints nevertheless return to the same 32-dimensional lookup-latent space and the same frozen cylinder decoder.
Structured physics adaptation
Frozen-backbone low-rank correction
After generator selection, both and , as well as the latent inverse transform, are frozen. For a decoder-derived adapter input feature , the parallel low-rank adapter[57, 58] has the generic form
where is the adapter rank, its scale, and an optional condition gate. For an input feature of dimension and a corrected head of dimension , and ; is a fixed branch-specific output scale. is initialized to zero so the learned branch begins with no correction. Data anchors, frozen-field teachers, trust regions, and parameter regularization restrict the correction to the neighborhood of the pretrained field manifold.
Cavity reduced-BGK adaptation
The cavity adapters act on the four standardized primitive outputs, the two 256-dimensional kinetic coefficient vectors, and the two scalar nonequilibrium biases. Their output scales are 0.10, 0.10, and 0.02, respectively. A width-32 condition network with Knudsen Fourier frequencies produces five sigmoid gates: one shared by the four primitive corrections, one for each of the - and -coefficient branches, and one for each of their scalar-bias branches. Because the corrections are inserted before Eq. (41), positivity and the bounded nonequilibrium parameterization are preserved.
The steady reduced-BGK residual is evaluated in logarithmic distribution form,
with from Eq. (3). Central differences use , and the residual is integrated over all 6,400 collocation nodes. Unless stated otherwise, training and checkpoint probes use a near-wall-biased collocation measure; a separate uniform-area sensitivity audit is reported in Section C. With normalized velocity weights , the reported matched-grid BGK diagnostic is
where the brackets average the sampled spatial locations and conditions. Additional terms measure diffuse-wall consistency, conservation of collision invariants, and agreement between kinetic moments and the macroscopic head.
For each component , eight independently sampled frozen-decoder probe batches define a detached scale
The normalized physical objective is
where identifies the frozen generator family, and for FM and for diffusion. Thus, conservation and moment consistency are protected by one-sided guards rather than minimized without limit. The floors are .
The DVM anchor loss is the equal mixture of losses evaluated with optimized lookup latents and generator-produced latents. Each constituent is the sum of a standardized primitive-field mean-square error and the two quadrature-weighted relative errors of and . At continuously sampled Knudsen numbers, a frozen-decoder teacher penalizes changes in the primitive outputs, , and the actual distributions under the same generated latent. The total objective is
where is the optimizer-update index and . The regularizer contains squared hinges beyond trust radii for the macroscopic RMS, relative coefficient RMS, and bias RMS, plus times the mean squared composed weight . The FM and diffusion adapters use ranks 8 and 16, respectively, reflecting their separately frozen model-selection protocols.
For reporting and checkpoint confirmation, the frozen and adapted models are evaluated on the same independently seeded probe batches. Let be Eq. (71) evaluated for on probe , using the same detached scales. The quantities reported in Section IV are
where an overbar denotes the mean over paired probes and . A “worst-probe” reduction is the minimum paired-probe reduction, whereas a worst-probe ratio is the maximum paired-probe ratio. These frozen-relative quantities are audit statistics, not additional optimization losses.
A reduction in Eq. (69) is interpreted only as improved consistency under the matched variables, quadrature, and differentiation rule. It neither proves satisfaction of the exact continuous BGK equation nor implies simultaneous improvement of every moment and boundary diagnostic.
Cylinder macroscopic adaptation
For the cylinder, the supplied to Eq. (67) concatenates the final frozen trunk feature in and generated lookup latent . A rank-16 branch corrects the four standardized macroscopic outputs, while a scalar branch supplies a localized normal-velocity correction:
where , , and, because the cylinder is centered at the origin, is the outward unit normal. The retained protocol also uses
which enforces zero normal velocity at the cylinder surface without imposing no-slip. Here and are the normal and tangential velocity components. This exact wall behavior is structural and must be distinguished from a residual reduction learned by the adapter. Each generator-specific cylinder adapter contains 9,893 trainable parameters.
Only constraints closed by the four available DSMC fields are used: wall no-penetration, upstream inlet consistency, local weak control-volume (CV) mass balance, and global boundary mass balance. For a fluid-only control volume ,
where is the outward unit normal to , and prevents division by zero. An analogous normalized flux is evaluated on the outer domain boundary. For a sampled mini-batch, define
where is the mean-square standardized four-field error for case , is the number of sampled cases, and is the complete adapter correction relative to the frozen decoder. The trust loss is the squared positive part of the ratio between correction RMS and frozen-field RMS plus 0.10, beyond a radius of 0.05; is the mean squared magnitude of the trainable adapter parameters. The complete objective is
For each raw physical loss , the detached normalization scale is initialized by its first observation and subsequently updated as
The component floors for wall, inlet, local CV, and global mass losses are , respectively. The half-cosine ramp is zero for the first 10% of training and reaches one at 30%. The FM adapter is exposed to optimized lookup latents; the diffusion adapter draws lookup and diffusion-generated latents with equal probability.
The scalar used for checkpoint selection and post-training reporting is a separate frozen-relative audit statistic, not the training term normalized by the exponential-moving-average (EMA) loss scales in Eq. (78). Let and denote the raw loss of component evaluated for the frozen and adapted predictions under the same audit samples and quadrature. The component ratio and aggregate physics score are
The weights match those in Eq. (78). Thus, indicates no aggregate change from the paired frozen generator, and indicates a weighted reduction. Because each generator has its own frozen denominator, this score measures relative improvement within a generator and is not an absolute residual for ranking FM against diffusion.
Momentum and total-energy residuals are not evaluated because the DSMC targets do not contain the pressure tensor, viscous stress, or heat flux required to close them. Complete BGK or Boltzmann residuals, Navier–Stokes constitutive terms, no-slip, gas-temperature-equals-wall-temperature, strong shock continuity, and symmetry penalties are likewise disabled.
Feasibility and checkpoint gates
A physics-adapted checkpoint is eligible only if it preserves the frozen decoded-field accuracy within the stated tolerance. Among eligible checkpoints, a normalized physics score ranks the correction. Cavity selection also guards kinetic accuracy, teacher consistency, collision invariants, moments, and the wall diagnostic. Cylinder selection constrains validation mean error to remain within a fixed feasibility margin of its frozen value; the retained diffusion protocol additionally applies the same margin to the worst case. A positive-step checkpoint is required, so a zero-initialized model cannot be reported as a trained physical improvement. These are multi-objective feasibility rules, not a claim that data and physics losses form a single universal scalar metric.
Training and evaluation protocols
Generator optimization settings
The retained conditional generators are summarized in Table III. AdamW[63], cosine learning-rate decay, gradient clipping at one, and an EMA of model weights are used in all generator fits, but the numerical precision and deployment solvers are model specific.
| Setting | Cavity FM | Cavity diffusion | Cylinder FM | Cylinder diffusion |
| Chart dimension | 4 | 4 | 12 | 12 |
| Parameters | 2.997 M | 2.997 M | 0.208 M | 0.203 M |
| Prediction/path | Trigonometric FM | VP v-prediction | Quadratic FM | VP x-prediction |
| Peak learning rate | ||||
| Batch size | 4096 | 4096 | 512 | 512 |
| Training updates | 30000 | 8000 | 8000 | 12000 |
| Precision | BF16/FP32 | FP32 | FP32 | FP32 |
| Solver | Heun-48 | DPM-Solver++(2M)-35 | Heun-24 | DDIM-40 |
| Ensemble/anchors | 64 | 64 | 3 anchors | 40 |
Table III. Frozen conditional-generator configurations. Ensemble counts refer to the validation protocol; larger counts are used for final cavity field evaluation.
In the cavity FM fit, BF16 is used for the principal regression and FP32 for the curvature term; all other entries use FP32 throughout. A notation such as Heun-48 or DDIM-40 specifies the solver followed by its number of integration steps.
Ablation and checkpoint selection
Generator ablations change one principal modeling choice at a time and rank checkpoints by held-out decoded-field error rather than by training loss. The cavity FM study compares the trigonometric FiLM reference, a rectified path, the capacity-matched plain-concatenation network, a smaller network, removal of Knudsen Fourier features, and removal of the Knudsen-curvature term. Separate follow-up controls test velocity-field ensembling and differentiable endpoint mean/variance regularization. The cavity diffusion study first compares -, -, and -prediction, BF16 and FP32 arithmetic, plain-concatenation and Knudsen-FiLM conditioning, and endpoint regularization. After fixing the selected denoiser, it scans and the DDPM, DDIM, and DPM-Solver++ sampler families, including the deployment step count.
For the cylinder, the initial six-way comparison varies the FM path, plain versus FiLM conditioning, and the three diffusion prediction targets. The subsequent FM study holds the local four-neighbor training graph fixed and then tests, in sequence, the rank-eight quadratic condition correction, simplex rather than distance aggregation, decoder-metric rather than ordinary PCA, decoded-field endpoint supervision, multi-source consensus, and optional composition/loop terms.
This staged design separates improvements in latent geometry from changes in the transport network and numerical aggregation. Strict leave-one-condition-out refits are reserved for the selected candidates; in each fold, all latent-chart and graph quantities are reconstructed without the omitted condition.
Physics-adapter ablations are performed only after the corresponding frozen generator is selected. The cavity study varies adapter rank and the normalized BGK, boundary, conservation, moment, anchor, and teacher components. The cylinder study compares a data-only adapter, wall/inlet terms, mass-balance terms, ranks eight and 16, their combined objective, and the optional hard wall-normal projection. Eligibility is always determined by the feasibility gates in § 3 F; a smaller physics residual cannot compensate for violation of the decoded-field accuracy guard. These experiments provide component-removal and coarse-strength evidence, but they are not an exhaustive continuous search over every scalar loss weight and ramp location. The reported weights should therefore be interpreted as one validated feasible setting rather than a uniquely identified optimum.
Physics-adapter training settings
The structured-adapter protocols are summarized in Table IV. All listed runs begin from a new zero-output adapter and do not resume an earlier adapter checkpoint.
| Setting | Cavity FM | Cavity diffusion | Cylinder FM | Cylinder diffusion |
| Adapter rank | 8 | 16 | 16 | 16 |
| Trainable parameters | 41,685 | 75,525 | 9,893 | 9,893 |
| Training updates | 1500 | 1500 | 12000 | 12000 |
| Maximum branch learning rate | ||||
| Latent exposure | Lookup/FM, 1:1 | Lookup/diffusion, 1:1 | Lookup only | Lookup/diffusion, 1:1 |
| Conditions or cases/update | 32 | 32 | 8 | 16 |
| Field-anchor points/case | 384 | 384 | 4096 | 8192 |
| Interior collocation points | 512 | 512 | – | – |
| Boundary quadrature count | 96/wall | 96/wall | 512 | 1024 |
| Velocity nodes (physics/anchor) | 6400/4096 | 6400/4096 | – | – |
| CVs face points | – | – |
Table IV. Physics-adapter configurations. The two cylinder columns share the same corrected outputs and physics definitions but differ in latent exposure and per-update sampling.
Here “lookup” denotes the optimized representation-stage code for an observed training case; it is not an Oracle code fitted to a held-out case.
Data separation and reproducibility
Architecture and checkpoint selection use only the development partitions defined in Section II. Additional out-of-sample cases do not affect normalization, latent-chart fitting, optimizer updates, or checkpoint ranking. During leave-one-condition-out studies, the omitted lookup code is also excluded from PCA or decoder-metric fitting, score normalization, graph construction, and training pairs. All aggregate errors first assign equal weight to each condition as in Eq. (31).
Reproducibility artifacts store the resolved configuration, random seed, normalization and chart state, selected model weights, upstream checkpoint paths, and cryptographic hashes. Field evaluation reloads the frozen artifacts rather than reconstructing architectures from manually transcribed hyperparameters. The reported ensemble size, numerical solver, and checkpoint must therefore be treated as part of each frozen model definition rather than as interchangeable post-processing choices. Training wall times and a controlled single-GPU inference benchmark are reported in Section E; because the distributed fields include no solver job records, no solver-to-surrogate acceleration ratio is reported. A 20-sampling-seed audit of the frozen cavity inference (Section D) shows sampling variability far below the method differences reported below; it does not quantify retraining uncertainty, which the five-seed refits of Section V A address separately.
Model selection and controlled ablation studies
This section fixes the representation, generator, and adapter configurations used in the out-of-sample study. All choices are made on the development partitions defined in Section II or by strict leave-one-condition-out (LOCO) refitting of the training set. The additional cases used in § 4 are not used for the comparisons below. The data partitions, reporting metrics, and equal-case aggregation were fixed before final checkpoint scoring. The architecture candidates were developed iteratively on the development data and should not be interpreted as an exhaustively preregistered search space. Retention therefore follows the stated multi-criterion protocol—decoded-field accuracy, endpoint stability, and gate feasibility—rather than selecting the minimum entry of each table. Cavity results are reported using the relative errors in Eqs. (26) and (27), whereas cylinder model selection uses the area-weighted standardized RMSE of the four decoder outputs. For condition , define
where indexes spatial samples, is the spatial quadrature weight, and contains the four standardized disturbances in Eq. (19). This selection metric differs from the dimensionless per-field RMSE in Eq. (30), which is used for physical-field reporting. Neither metric is numerically comparable with the cavity relative errors. Throughout this section, boldface in ablation and confirmation tables denotes a configuration retained for later evaluation, not necessarily the minimum in every reported column.
An Oracle result freezes all decoder parameters and optimizes only one latent code directly against the complete held-out reference field. It is a representation-floor diagnostic, not a deployable conditional prediction; the converged cavity and cylinder values below use 200 and 6,000 latent-only updates, respectively. In a LOCO fold, condition and its lookup code are removed from every fitted chart and conditional model before predicting that condition. For any per-condition error , the reported aggregates are
The validation mean and worst columns use the same mean and maximum operations on the fixed validation partition, without refitting. Unless stated otherwise, a reported reduction of a nonnegative diagnostic is , and a signed percentage change is .
For generated chart endpoints at condition , define
The “spread range” is over the stated cases. “Training spread” is the root mean square of the componentwise population standard deviations over all observed training conditions and chart coordinates, equivalently the aggregate form of Eq. (83). Both are evaluated in the transformed chart and diagnose numerical endpoint dispersion rather than physical uncertainty.
Representation fidelity and latent-chart dimension
The cavity auto-decoder attains an Oracle validation mean of and for . In contrast, direct linear interpolation of the 24-dimensional lookup codes in log Kn gives . The frozen decoder can therefore represent the held-out solutions more accurately than a simple condition-space interpolation can locate them. The gap between these two errors is attributable to latent inference under a fixed decoder, although the nonzero Oracle error remains part of the total prediction error.
The retained cavity chart in Eq. (47) is selected by decoding each truncated latent rather than by explained variance alone. If is the inverse-transformed rank- reconstruction of lookup code , define
The mean and worst increments in Table V are the mean and maximum of over the training cases. For each rank, the LOCO column comes from the same fixed-capacity conditional MLP refitted after excluding the held-out lookup code; this MLP is used only as a train-only chart-selection probe and is not one of the deployed generators. As shown in Table V, rank four reduces the worst truncation-induced kinetic increment to and gives the smallest decoded LOCO error. Rank eight has a slightly smaller reconstruction increment but a larger LOCO error, demonstrating that additional weak latent directions can make conditional learning harder without materially improving field representation. The small negative mean increment at rank four is a finite-precision cancellation effect.
| PCA rank | Explained variance | Mean increment | Worst increment | LOCO |
| 2 | 99.5975% | |||
| 4 | 99.9958% | |||
| 8 | 99.9999% |
Table V. Cavity latent-chart ablation under decoded LOCO evaluation.
Because the chart is estimated from only 16 lookup codes, a leave-one-condition-out principal-angle audit (Section B) additionally shows that the dominant four-dimensional span is stable for most omissions (mean largest angle ) but not at the training-support boundary (maximum ), limiting claims about an intrinsic four-dimensional physical manifold.
For the cylinder, converged frozen-decoder Oracle optimization gives final mean and worst values of equal to 0.01521 and 0.02083, respectively; restricting the same metric to cells with gives a mean near-wall value of 0.05370. A common chart dimension of 12 is fixed before generator training by complete train-only LOCO decoded-field screening. The ordinary unwhitened PCA-12 chart used by the diffusion baseline retains 95.51% of the training-latent variance. The retained FM instead uses the separate decoder-metric PCA-12 chart in Eq. (48); its dimension is the same, but its metric and coordinate scaling are not. Whitening is rejected because it gives weak, poorly predictable latent directions the same regression weight as dominant directions. Condition Fourier features are rejected separately because they reduce training error while degrading omitted-condition field prediction. The nonzero Oracle floor, especially near the cylinder, is retained in all subsequent error accounting rather than attributed to the conditional generator.
Flow-matching ablations
Cavity: conditioning and probability path
The six cavity FM candidates use the path and objective in Eqs. (52) and (55), the same rank-four chart, training cases, 30,000-update budget, and decoded-field evaluation. The FiLM reference and the plain-concatenation candidate contain 2,970 and 2,997 million parameters, respectively, so their comparison is effectively capacity matched. Table VI shows that the simpler plain concatenation of the state, time, and Knudsen features is substantially more accurate than repeated FiLM conditioning. Removing Knudsen Fourier features also increases endpoint spread, while the original curvature penalty is not beneficial for the FiLM reference. The trigonometric FiLM reference has width 256 and six residual blocks. The candidate labeled linear optimal transport (linear-OT) replaces only its probability path by ; the reduced FiLM uses width 128 and three blocks. The remaining two controls remove, respectively, the Knudsen Fourier lifting and Eq. (54).
| Candidate | Mean | Maximum | Spread range |
| Trigonometric FiLM | 0.0175–0.0761 | ||
| Linear-OT FiLM | 0.0109–0.0667 | ||
| Plain concatenation | 0.0145–0.0210 | ||
| Reduced FiLM (width 128, 3 blocks) | 0.0158–0.0831 | ||
| No Knudsen Fourier features | 0.0515–0.1507 | ||
| No curvature penalty | 0.0107–0.0335 |
Table VI. Cavity FM architecture ablation on three held-out development Knudsen numbers; spread is measured in the transformed rank-four chart.
Plain concatenation lowers the mean error by 70.8% relative to the trigonometric FiLM reference. A plausible mechanism is that repeated feature-wise modulation gives the sparsely sampled condition signal multiple opportunities to rescale hidden activations, so small condition-fitting errors can be amplified across residual blocks. Plain concatenation instead exposes the state, time, and condition features directly to one smooth vector field. This is an interpretation of the controlled ablation, not a general claim that FiLM is inferior in larger or more densely sampled data sets. Increasing the antithetic ensemble from 64 to 4096 changes its mean only from to , so the residual is not primarily Monte Carlo error. A four-member velocity-field ensemble instead gives ; this control trains four independently initialized vector fields and averages their velocities at every ODE evaluation. Differentiable endpoint mean matching and variance contraction, respectively penalizing and excess component variance above , also increase the mean to ; they correct statistics at observed Knudsen numbers but do not supply the missing mean at an unobserved condition. These controls support freezing the single plain concatenation FM without endpoint regularization.
Cylinder: decoder-aware curved transport
The cylinder study uses strict LOCO refits because ordinary validation alone can favor a model that interpolates the particular split. Starting from the local source-to-target FM in Eqs. (57) and (62), the ablation successively adds a rank-eight quadratic condition correction, simplex aggregation, decoder-metric latent coordinates, decoded-field supervision, source consensus, and loop regularization. The selected model is the simplest candidate containing the decoder metric, field term, and consensus term; the loop penalty is rejected. Here distance aggregation uses the nearest anchors with positive distance-softmax weights, whereas simplex aggregation uses the barycentric weights of a local enclosing triangle when one exists. The loop control adds to penalize round-trip inconsistency.
Relative to the best preceding first-order transport, the selected curved model reduces the LOCO mean and worst errors by 16.66% and 22.22%. Its largest remaining error occurs at the parameter-domain boundary case and is concentrated in the shock and near-wall regions. The fact that simpler candidates have smaller ordinary validation errors in Table VII is not used to reverse the selection: the architecture was fixed by the more stringent omitted-case protocol.
| Candidate | LOCO mean | LOCO worst | Validation mean | Validation worst |
| Quadratic PCA, distance anchors | 0.05835 | 0.27137 | 0.03584 | 0.13565 |
| Quadratic PCA, simplex | 0.05608 | 0.26740 | 0.03825 | 0.15473 |
| Quadratic metric, simplex | 0.05301 | 0.22635 | 0.04003 | 0.18258 |
| Metric, field, consensus loop | 0.05222 | 0.20741 | 0.04054 | 0.18709 |
| Metric, field, consensus | 0.05114 | 0.20322 | 0.04095 | 0.18766 |
Table VII. Cylinder curved-transport ablation under the LOCO criterion in Eq. (81).
Diffusion ablations
Cavity: target, precision, schedule, and sampler
The most consequential cavity diffusion choices are numerical precision and the regression targets defined in Eq. (64). As summarized in Table VIII, changing only BF16 to FP32 lowers the -prediction mean by 37.5%. In FP32, -prediction is markedly more accurate than clean-latent () prediction or stabilized noise () prediction. Native -prediction is numerically ill-conditioned near the high-noise endpoint and is not retained. The stabilized control truncates diffusion time at and applies minimum signal-to-noise-ratio (SNR) weighting with , i.e., weight for the -prediction mean-squared error (MSE). The native control uses the unweighted objective and .
| Target and arithmetic | Mean | Maximum | Training spread |
| , BF16 | – | ||
| , FP32 | |||
| , FP32 | |||
| , FP32, , Min-SNR | |||
| , FP32, unweighted |
Table VIII. Cavity diffusion precision and prediction-target ablation with a common decoder and rank-four chart. Training spread is the RMS endpoint spread in the transformed chart.
Endpoint mean and variance losses change the FP32 -prediction mean from to and , respectively, and Knudsen-FiLM conditioning gives . None improves on plain input concatenation. An 8,000-update model is also retained over 20,000 updates because the latter reduces training spread but increases validation by 1.17%.
After fixing FP32 -prediction, the development scan of Eq. (63) identifies as an empirical low-error range and selects . Under the 8,000-update fit, changing the cosine VP schedule , to the continuous linear- VP schedule lowers the DDIM-48 kinetic mean from to . The subsequent sampler study is summarized in Table IX. DPM-Solver++(2M)-35 is frozen because the primary selection metric is the mean error; the 44-step alternative is retained only as a worst-case diagnostic.
| Sampler | Steps | Mean | Maximum |
| DDIM | 48 | ||
| DPM-Solver++(2M) | 35 | ||
| DPM-Solver++(2M) | 44 |
Table IX. Cavity diffusion sampler comparison over 20 paired random seeds. Uncertainty is the population standard deviation, and the last column is the seed–condition maximum.
Both DPM-Solver++ configurations outperform DDIM in all 20 paired seeds. The seed variation measures numerical Monte Carlo stability of the ensemble mean; it does not imply multiple physical steady states at a fixed Knudsen number. The 37.5% BF16-to-FP32 gap in Table VIII is also a practical result: low-dimensional latent diffusion is inexpensive enough that full-precision training is preferable here, and mixed precision should not be assumed benign merely because the decoded fields are smooth.
Cylinder: Gaussian-reference diffusion baseline
The cylinder target ablation produces a different ordering. With identical rank-12 charts and 12 held-out training conditions, -prediction gives the smallest LOCO mean, -prediction the smallest LOCO worst error, and -prediction is inferior on both metrics:
| Target | LOCO mean | LOCO worst | Validation mean | Validation worst |
| x-prediction | 0.15537 | 0.59472 | 0.05326 | 0.17129 |
| -prediction | 0.17196 | 0.48652 | 0.03809 | 0.09293 |
| -prediction | 0.24199 | 0.63904 | 0.04441 | 0.11545 |
Table X. Cylinder Gaussian-reference diffusion target ablation in the train-only screen using the preregistered LOCO-mean criterion.
The predeclared train-only mean criterion selects -prediction without consulting the validation ordering. In the complete 48-fold confirmation it obtains LOCO mean/worst errors of 0.07284/0.46326, compared with 0.03896/0.16429 for the contemporaneous trigonometric FM. This comparison belongs to the initial Gaussian-reference screen; the subsequently retained curved FM is selected separately in § 3.2 B. No further diffusion schedule or sampler search is conducted because the candidate fails the preregistered promotion criterion. The retained cylinder diffusion comparison model, linear VP with and DDIM-40, serves as a controlled generative comparison rather than the principal accuracy model.
Physics-adapter ablations
Cavity reduced-BGK correction
For cavity FM, the unnormalized physics objective reduces the matched-grid BGK diagnostic by 47.4% but increases kinetic error by 16.0% and approximately doubles the conservation and moment diagnostics. The normalized, manifold-anchored objective in Eqs. (71) and (72) instead retains the frozen prediction accuracy while reducing BGK by 28.65%; its principal results are compared with the formal diffusion adapter in Table XI. Here the BGK and composite-physics reductions are and from Eq. (73); conservation and moment values are the corresponding frozen-relative ratios .
| Model | Frozen | Adapted | BGK reduction | Physics reduction |
| FM, unnormalized objective | 47.4% | – | ||
| FM, normalized manifold anchor | 28.65% | 22.06% | ||
| Diffusion, rank 16 | 23.59% | 19.35% |
Table XI. Cavity physics-adapter selection on held-out development cases. Reductions are relative to each frozen generator and are not compared across generator families.
The normalized FM adapter changes the kinetic mean by , but its independent-probe conservation and moment ratios are both 1.0574. The checkpoint is therefore retained as Physics-constrained FM. This terminology denotes a reduction of the registered composite and BGK diagnostics subject to bounded invariant degradation; it does not mean that the invariant diagnostics themselves improve. A subsequent positivity-preserving hard-moment projection reduces its internal moment defect below , where the defect is the RMS relative mismatch between the projected kinetic moments and their macroscopic-head targets. It also increases kinetic error and the pressure-tensor error in Eq. (28) by 4.08% and 30.10%, respectively; this hard-projection variant fails the joint gates and is rejected.
For diffusion, the controlled rank/weight/anchor ablation selects rank 16. During selection it changes from to , while the mean/worst composite-physics reductions are 19.21%/15.42%. Rank four produces a larger 23.29% mean physics reduction but increases kinetic error by 1.49%; rank eight gives a 21.08% reduction and a 3.17% kinetic improvement. The independent-seed rank-16 confirmation in Table XI passes every registered gate, including worst-probe conservation and moment ratios of 0.9331. Removing the invariant guards or using physics weights of and fails the joint physical-reduction gates, confirming that physical loss magnitude alone is not a sufficient checkpoint criterion. A robustness audit (Section F) repeats this adapter with two additional training seeds and with the physics total weight scaled by and under a common sampling protocol: the seed variants span BGK reductions of 24.5–32.6% with invariant ratios 1.027–1.069, the weight variants yield smaller reductions of 10.2–14.2% with invariant ratios of 1.014–1.045, and field accuracy is preserved in every run, so the adapter conclusion is stable to seed and weight perturbations.
An independent frozen-checkpoint audit of the differentiation step and query-grid refinement (Section C) shows that the BGK reduction is stable to the difference step (0.223 and 0.104 percentage-point spans for FM and diffusion) while its magnitude depends on the spatial sampling measure; the near-wall-biased values in Table XI should not be read as uniform-area residual reductions.
Cylinder boundary and weak-mass constraints
The cylinder screen isolates adapter capacity from the individual evaluable physics terms in Eqs. (75) and (76). Its complete result is given in Table XII. The mass-only candidate returns to the zero-output checkpoint because its trained checkpoints do not improve the independent composite audit. Rank eight and rank 16 are nearly tied without the hard map; the rank-16 hard-wall candidate is retained because it gives the lowest feasible physics score without consuming the field-error budget.
| Candidate | Physics score | ||
| Frozen/data-only control | 0.04095 | 0.18766 | 1.00000 |
| Wall and inlet, rank 16 | 0.04105 | 0.18755 | 0.88803 |
| Mass balance only, rank 16 | 0.04095 | 0.18766 | 1.00000 |
| Joint boundary–mass objective, rank 8 | 0.04093 | 0.18747 | 0.90959 |
| Joint boundary–mass objective, rank 16 | 0.04099 | 0.18748 | 0.90672 |
| Joint objective + hard wall, rank 16 | 0.04094 | 0.18737 | 0.69893 |
Table XII. Cylinder FM-adapter screen using the frozen-relative physics score in Eq. (80).
The 12,000-update confirmations are reported in Table XIII. The FM adapter changes the exhaustive mean/worst field errors by +0.094%/−0.029%; its wall, inlet, local-CV, and global-mass ratios are , 0.2769, 1.0218, and 0.9630. A paired counterfactual audit in Table XIV, evaluated on an independent fixed sample set, separates inference-time use of the learned correction and analytic hard map while holding the decoder, generated latents, adapter weights, quadrature, and audit seeds fixed. The hard map alone already drives the wall ratio to with the remaining ratios unchanged; the jointly deployed model additionally reduces the inlet ratio to 0.2769 and the global-mass ratio to 0.9629, while the local-CV diagnostic does not improve. The diffusion adapter improves its mean error by 1.63%, increases the worst error by 2.62%, and reduces all four registered physical diagnostics, with ratios 5.14 × 10⁻¹¹, 0.4885, 0.9174, and 0.9602. Here “exhaustive” means that Eq. (81) is evaluated over all available native-mesh cells in every validation case rather than an audit subsample.
| Generator | Frozen mean | Adapted mean | Frozen worst | Adapted worst | Physics score |
| FM | 0.04260 | 0.04264 | 0.19955 | 0.19949 | 0.68011 |
| Diffusion | 0.05866 | 0.05770 | 0.23020 | 0.23623 | 0.65369 |
Table XIII. Formal cylinder adapter confirmation. Field-error columns use Eq. (81); the physics score in Eq. (80) is evaluated independently for each generator.
| Inference variant | Wall | Inlet | Local CV | Global | ||
| Frozen decoder | 0.04260 | 1.0000 | 1.0000 | 1.0000 | 1.0000 | 1.0000 |
| Hard map only | 0.04263 | 1.0000 | 1.0035 | 1.0000 | 0.7516 | |
| Learned adapter only | 0.04273 | 58.9336 | 0.2769 | 1.1430 | 0.9629 | 15.4680 |
| Learned adapter + hard map | 0.04264 | 0.2769 | 1.0219 | 0.9629 | 0.6801 |
Table XIV. Paired attribution of the formal cylinder FM correction. All rows use the same checkpoint and audit samples. Component entries are frozen-relative ratios from Eq. (80); is the exhaustive mean field error.
The lower relative diffusion physics score means that its adapter produces a slightly larger fractional change from its own baseline. It does not mean that diffusion has smaller absolute residuals. Under the common exhaustive field audit, the Physics-constrained Diffusion model has 35.31% higher mean error than Physics-constrained FM, which therefore remains the principal cylinder model.
The learned-only row is a post-training counterfactual, not a separately trained no-map model. Its wall ratio of 58.93 shows that the learned correction and analytic projection are not additive modules: during joint training, the hard map removes wall-normal velocity after the adapter, so the adapter alone is not constrained to remain feasible at the wall. Consequently, neither the joint reduction nor its difference from the hard-only case should be attributed entirely to learning. The defensible conclusion is narrower: the analytic map supplies exact no-penetration, while the learned correction adds inlet and global-balance improvements only when deployed with that map.
Because the two pipelines differ in latent chart, reference state, latent exposure, and per-update sampling, this result ranks the retained pipelines rather than isolating the generator algorithm; the controlled differences are summarized in § 3 E and Table IV. Neither result establishes the complete Boltzmann equation: only wall no-penetration, inlet consistency, and weak local/global mass balance are closed by the available cylinder variables.
Frozen configurations for out-of-sample evaluation
The ablations above define the four configurations in Table XV. No architecture, sampler, ensemble size, or adapter checkpoint is changed after examining the additional out-of-sample cases in § 4.
| Problem | Generator | Latent chart | Conditional dynamics | Deployment | Physics adapter |
| Cavity | FM | whitened PCA-4 | plain-concat FM | trigonometric Heun-48, 256 samples | rank 8, normalized manifold BGK |
| Cavity | Diffusion | whitened PCA-4 | FP32 VP -prediction | DPM-Solver++(2M)-35, 256 samples | rank 16, normalized BGK |
| Cylinder | FM | decoder-metric PCA-12 | quadratic local transport | anchor Heun-24, 3 anchors, 21 | rank 16, hard wall |
| Cylinder | Diffusion | unwhitened PCA-12 | VP -prediction, | DDIM-40, 40 samples | rank 16, hard wall |
Table XV. Final frozen conditional generative configurations. The cylinder diffusion branch is retained as a controlled comparison, whereas the curved FM branch is the principal cylinder accuracy model.
In summary, the retained configurations are not selected by training loss alone. Cavity FM favors a simpler conditional vector field, cavity diffusion requires FP32 -prediction and a higher-order deterministic sampler, and cylinder FM benefits from local decoder-aware curved transport. The reported adapters pass their benchmark-specific final acceptance rules; for cavity FM, the retained checkpoint satisfies the invariant gate. This separation permits § 4 to evaluate generalization without further model selection. The deterministic baselines in Eq. (36) are frozen under the same data-separation rule and are evaluated only after generator selection. Because that comparison uses additional conditions, its decoded-field results are reported in § 4 A, rather than used as another development-set ablation.
Out-of-sample predictions and discussion
The configurations frozen in § 3.2 E are evaluated without further fitting or checkpoint selection. They remain accurate over a range of out-of-sample conditions; the following discussion focuses on the representative cases declared in Section II. Throughout this section, Frozen FM and Frozen Diffusion denote the unadapted conditional generators. Physics-constrained FM and Physics-constrained Diffusion denote the corresponding frozen generators followed by their accepted structured adapters. Neither the generator nor the latent transform is refitted on these additional cases. Conditional generative surrogates for partial differential equations learn a condition-dependent distribution of solution fields[47, 48]; freezing every learned component here ensures that the following results measure that conditional generalization rather than test-time adaptation. Sampling can reduce numerical Monte Carlo variability, but it cannot add missing condition-space evidence.
Deterministic baseline under frozen decoded-field evaluation
To determine whether conditional generation improves steady point prediction, Table XVI compares the direct MLP in Eq. (36) with the two frozen generative pipelines. No predictor, chart, decoder, or normalization statistic is refitted on these cases. For the cavity, all three predictors share the whitened PCA-4 chart and the same decoder. For the cylinder, the MLP and diffusion share the ordinary PCA-12 chart, whereas FM retains the decoder-metric PCA-12 chart selected in § 3.2; the latter is therefore a pipeline-level comparison. Each MLP entry reports the mean over five independent training seeds with the population standard deviation in parentheses. Two capacity tiers are trained on every chart: the compact ensembles (203,532 cavity and 89,340 cylinder parameters) and single-member parameter-matched controls with 2,989,188 (cavity) and 209,772 (cylinder) parameters, matching the retained generators to within about 3%. The cylinder MLPs are evaluated on both the ordinary and the decoder-metric PCA-12 charts, so the deterministic baselines bracket the chart choices of the two generators. For the cylinder rows, define the descriptive six-field average
| Problem/subset | Predictor | Chart | Parameters | Mean | Worst case |
| Cavity, interpolation<br>Kn = 0.23, 0.73 | Deterministic MLP (compact) | whitened PCA-4 | 203 532 | ||
| Cavity, interpolation<br>Kn = 0.23, 0.73 | Deterministic MLP (matched) | whitened PCA-4 | 2 989 188 | ||
| Cavity, interpolation<br>Kn = 0.23, 0.73 | Frozen FM | whitened PCA-4 | 2 996 932 | ||
| Cavity, interpolation<br>Kn = 0.23, 0.73 | Frozen Diffusion | whitened PCA-4 | 2 996 932 | ||
| Cavity, extrapolation<br>Kn = 0.03, 1.00 | Deterministic MLP (compact) | whitened PCA-4 | 203 532 | ||
| Cavity, extrapolation<br>Kn = 0.03, 1.00 | Deterministic MLP (matched) | whitened PCA-4 | 2 989 188 | ||
| Cavity, extrapolation<br>Kn = 0.03, 1.00 | Frozen FM | whitened PCA-4 | 2 996 932 | ||
| Cavity, extrapolation<br>Kn = 0.03, 1.00 | Frozen Diffusion | whitened PCA-4 | 2 996 932 | ||
| Cylinder, three shown | Deterministic MLP (compact) | ordinary PCA-12 | 89 340 | ||
| Cylinder, three shown | Deterministic MLP (matched) | ordinary PCA-12 | 209 772 | ||
| Cylinder, three shown | Deterministic MLP (compact) | decoder-metric PCA-12 | 89 340 | ||
| Cylinder, three shown | Deterministic MLP (matched) | decoder-metric PCA-12 | 209 772 | ||
| Cylinder, three shown | Frozen FM | decoder-metric PCA-12 | 208 144 | ||
| Cylinder, three shown | Frozen Diffusion | ordinary PCA-12 | 203 404 |
Table XVI. Frozen deterministic and generative predictors under common decoded-field evaluation. Cavity errors are from Eq. (27); cylinder errors are from Eq. (85). MLP entries are five-seed means with population standard deviations in parentheses; compact and parameter-matched capacity tiers are listed separately. Cylinder entries are evaluated over the three displayed conditions, not the complete additional evaluation set. Bold entries are the minimum mean within each problem and subset.
where every component is the dimensionless field RMSE from Eq. (30). This equal weighting is a compact reporting statistic, not a thermodynamic norm.
Repeating the generator fits with five independent seeds places the single-fit rows in context: the cavity FM and diffusion interpolation means, and , are statistically indistinguishable and remain above the MLP means (–); on extrapolation the diffusion mean is level with the MLPs and clearly better than FM, . On the cylinder the FM five-seed mean , against for diffusion, confirms a seed-robust pipeline ranking, while every MLP variant (seven-condition means – over the complete additional evaluation set; the cylinder rows of Table XVI report only the three displayed conditions) outperforms both generators. Thus, conditional generation does not provide a universal point-accuracy advantage for these single-reference steady problems. Its role here is to model cross-condition evolution as transport on a compact solution manifold and to retain a sampling-based inference mechanism on which the frozen physics-adaptation protocol can operate. The deterministic baseline also shows that this machinery is not required merely to obtain a strong conditional mean: for a single-valued steady map, a much smaller regressor can provide equal or lower decoded-field error. The generative formulation becomes more distinctive when a condition admits multiple states or irreducible variability, because it can in principle represent a conditional distribution instead of only its point estimate. That extension is architectural rather than empirical evidence from the present datasets, which contain one steady reference per condition. Accordingly, the current endpoint spread remains a transport-sensitivity diagnostic, not calibrated uncertainty, and the results do not show that the additional sampling machinery dominates a compact deterministic surrogate.
Lid-driven cavity flow
Interpolation within the training support
The conditions and probe distinct locations inside the training support. Figure 2 to Figure 5 compare the DVM fields with Frozen FM, Physics-constrained FM, Frozen Diffusion, and Physics-constrained Diffusion. At both conditions, the density, velocity, speed, and temperature contours are visually coincident at the plotted scale. The direct-error panels localize the remaining discrepancy near the moving lid, the upper corners, and the recirculating core rather than revealing a generator-specific large-scale distortion. The change from to is nevertheless resolved: increasing rarefaction broadens the lid-driven response and reduces the penetration of the high-speed layer into the cavity. The concentration of error near the lid–sidewall junctions is consistent with the simultaneous presence of a discontinuous wall velocity and a kinetic boundary layer, both of which make the local solution more sensitive to spatial and velocity-space resolution10,29.

Figure 2. Dimensionless cavity fields at from DVM and the four generative models. Columnwise color limits are shared across rows.
Figure 5. Direct dimensionless errors at . Each column uses 99th-percentile upper clipping.
At , the four macroscopic errors lie between and , while the kinetic errors lie between and . The FM and diffusion physics adapters reduce their matched-grid BGK diagnostics from to and from to , respectively, without a material change in field error. At , the four macroscopic errors lie between and and the kinetic errors between and ; the discrepancy is again concentrated near the moving lid and upper corners but spreads more deeply into the core, consistent with the thicker lid-driven kinetic layer of the more rarefied flow. Diffusion gives the smaller macroscopic error, whereas FM gives the smaller kinetic error. Physics correction reduces the FM kinetic error from to and the diffusion error from to ; the corresponding BGK diagnostics decrease from to and from to . Thus, for both displayed interpolation conditions, both adapted models improve the registered discrete BGK diagnostic while retaining the accuracy of the frozen fields. The FM invariant diagnostics increase by about , consistent with the selection-time audit in Section IV D; the diffusion invariant diagnostics decrease. This does not establish exact satisfaction of the continuous BGK equation: the diagnostic inherits the finite spatial differentiation and velocity quadrature of the matched DVM grid[59] [2].
As a pointwise rarefaction check, Figure 6 reports the mean gas slip and temperature jump in the central half of the first cell-center row below the moving lid. The DVM slip increases monotonically from at to at , while rises from , reaches approximately near , and then decreases slightly. The frozen and physics-constrained models reproduce both trends at , , and ; their visible discrepancy at is consistent with the low-side latent extrapolation error discussed below. Because these quantities are sampled at rather than extrapolated to the material wall, they are wall-adjacent diagnostics and are not presented as asymptotic slip coefficients. Their trend provides a physical check of the resolved rarefaction response without imposing no-slip or zero-temperature-jump behavior10,29.
Figure 6. Wall-adjacent moving-lid slip and temperature jump. DVM markers span the available conditions; model curves use the additional evaluation cases.
Extrapolation beyond the training support
The lower-side case is shown in Figure 7 and Figure 8. Its errors remain spatially structured and are much larger than those of the interpolation cases. Frozen FM and Frozen Diffusion macroscopic errors are and , respectively, with kinetic errors of and . The structured adapters change these values by less than in absolute terms. Thus, the dominant limitation lies upstream in the extrapolated chart endpoint; a capacity-limited local decoder correction cannot replace missing low-Knudsen condition coverage. The physical fields remain smooth, but the error maps show coherent rather than sample-like deviations across the domain. This supports the latent-endpoint diagnosis: the model produces a plausible cavity state, but one associated with the wrong position on the rapidly varying low-Knudsen solution chart. Upper-side extrapolation behaves differently. At , Figure 9 and Figure 10 show that both model families retain the global cavity structure. Diffusion gives macroscopic and kinetic errors of and , whereas FM gives and . The diffusion adapter lowers the kinetic error to and the matched-grid BGK diagnostic by 8.25%; the FM physics adapter lowers the BGK diagnostic by 12.20% but leaves its kinetic error essentially unchanged. The marked asymmetry between lower- and upper-side extrapolation shows that distance from the training interval alone does not determine difficulty; the regularity and continuation of the learned solution chart also matter. In particular, upper-side continuation preserves the broad spatial structures already present at the largest training Knudsen numbers, whereas lower-side continuation moves toward thinner wall layers and sharper gradients. The comparison therefore concerns the direction of continuation in solution space, not merely the scalar distance in .
Figure 7. Dimensionless cavity fields for lower-side extrapolation at . Columnwise color limits are shared across rows.
Figure 8. Direct dimensionless errors for lower-side cavity extrapolation at , using 99th-percentile upper clipping.
Figure 9. Dimensionless cavity fields for upper-side extrapolation at . Columnwise color limits are shared across rows.
Figure 10. Direct dimensionless errors for upper-side cavity extrapolation at , using 99th-percentile upper clipping.

Figure 3. Direct dimensionless errors at . Each column uses 99th-percentile upper clipping.
Rarefied cylinder flow
Aggregate results over additional conditions
The additional cylinder cases lie within the two-dimensional condition domain described in Section II B. The descriptive statistic is defined in Eq. (85); model selection remains that of § 3.2.
Table XVII shows that FM is more accurate than diffusion on this benchmark. Frozen FM gives , compared with 0.04748 for Frozen Diffusion, a 29.9% increase relative to FM. Five independent refits of each generator yield (FM) and (diffusion), so this ranking is robust to training seed. Pressure is the largest-error field for every model, followed by temperature and density, whereas both velocity components remain below in RMSE. Unlike the four decoder outputs, pressure is reconstructed as ; its error therefore combines correlated density and temperature errors and should not be interpreted as the error of a separately trained output head. Moreover, gives every reported field equal numerical weight, so it is a compact comparison statistic rather than a thermodynamic norm. Physics-constrained FM changes the equal-case mean from 0.03655 to 0.03670, while Physics-constrained Diffusion changes it from 0.04748 to 0.05008. Thus, neither physics-constrained model produces an aggregate field-accuracy
| Model | |||||||
| Frozen FM | 0.03769 | 0.11649 | 0.04089 | 0.00951 | 0.00571 | 0.00901 | 0.03655 |
| Physics-constrained FM | 0.03852 | 0.11636 | 0.04106 | 0.00952 | 0.00571 | 0.00901 | 0.03670 |
| Frozen Diffusion | 0.04235 | 0.15207 | 0.06003 | 0.01207 | 0.00684 | 0.01149 | 0.04748 |
| Physics-constrained Diffusion | 0.04518 | 0.15968 | 0.06205 | 0.01351 | 0.00715 | 0.01293 | 0.05008 |
Table XVII. Equal-case mean RMSE over the additional cylinder conditions. is the descriptive six-field average in Eq. (85).

Figure 4. Dimensionless cavity fields at from DVM and the four generative models. Columnwise color limits are shared across rows.
gain on these additional conditions.
This out-of-sample evaluation measures field error only. The DSMC output does not contain the higher-order fluxes required for a closed momentum, energy, BGK, or Boltzmann residual, and the common-seed physical audit used for checkpoint selection in § 3.2 D was not repeated on these additional conditions. Consequently, Table XVII neither confirms nor refutes whether the reductions in wall, inlet, and weak-mass diagnostics observed during physics-adapter selection transfer to these additional conditions. Moreover, the tabulated RMSE contains an unknown DSMC sampling-noise floor because independent realizations or averaging blocks are not available. The field-accuracy guard consequently limits degradation relative to this sampled reference; it is not a bound relative to a noise-free continuum mean.
Representative condition-dependent behavior
The representative cases are discussed in ascending Knudsen-number order. Because their Mach numbers also differ, the sequence illustrates the joint condition dependence and is not a one-parameter rarefaction sweep. At the lowest displayed condition, , Figure 11 and Figure 12, the flow is closest to the slip regime: the bow shock is thin and stands close to the cylinder, the stagnation compression is sharp, and the wake is comparatively narrow. The change from Frozen FM to Physics-constrained FM is negligible ( to ), whereas the change from Frozen Diffusion to Physics-constrained Diffusion increases the error from to . The degradation occurs across density, pressure, temperature, and velocity rather than in one isolated scalar. This case demonstrates that passing the frozen-field accuracy guard during adapter selection does not guarantee a field-error improvement at every additional condition.
Figure 11. Dimensionless cylinder fields at . Columnwise color limits are shared across rows and set independently for this condition. All cylinder panels in Figure 10–Figure 15 show the window , of the domain defined in Eq. (15), with the cylinder centered at the origin.

Figure 12. Direct dimensionless cylinder errors at , using 99th-percentile upper clipping.
At the intermediate condition , Figure 13 and Figure 14, the higher Mach number strengthens the compression and elongates the wake, while already thickens the bow shock into a broader compression layer and enlarges the near-wall Knudsen layer relative to . Physics-constrained FM reduces from to , while Physics-constrained Diffusion is effectively unchanged ( to ). The shock, wall, and wake structures are preserved after adaptation, consistent with the trust-region and manifold anchors restricting the correction to a small neighborhood of the frozen prediction.

Figure 13. Dimensionless cylinder fields at . Columnwise color limits are shared across rows and set independently for this condition.

Figure 14. Direct dimensionless cylinder errors at , using 99th-percentile upper clipping.
At the highest displayed Knudsen number, , Figure 15 and Figure 16, these trends are most pronounced: the shock appears as a broad smeared compression layer, the near-wall Knudsen layer is large, the wake is diffuse with weak recirculation, and the higher Mach number yields the strongest stagnation heating. Both adapters improve their own : FM changes from 0.03756 to 0.03737 and diffusion from 0.04812 to 0.04696. The residuals remain concentrated along the detached bow shock, in the near-wall layer, and through the downstream wake. Their location is physically consistent with the strongest condition-dependent structures in rarefied cylinder flow29.

Figure 15. Dimensionless cylinder fields at . Columnwise color limits are shared across rows and set independently for this condition.

Figure 16. Direct dimensionless cylinder errors at , using 99th-percentile upper clipping.
Across the three cases, the learned fields recover the principal shock and wake morphology of the statistically sampled DSMC references. Increasing rarefaction generally broadens kinetic structures, but the simultaneous Mach variation changes shock strength and prevents attribution of every visual difference to alone29. Fine-grained texture in the direct-error maps partly reflects the finite-particle statistical fluctuations intrinsic to DSMC averaging1; the coherent shock-aligned and wall-localized bands, by contrast, indicate systematic condition-dependent surrogate error. These two contributions cannot be separated quantitatively from a single time-averaged DSMC realization. Within each cylinder figure, color limits are shared across model rows but are set independently for each condition; absolute colors should therefore not be compared between different Knudsen–Mach pairs without reading their colorbars.
Cross-benchmark discussion
The two benchmarks do not support a universal ranking of conditional generators. Among the four displayed cavity conditions, diffusion has a clear advantage at and , while the two generators are closer and exchange the macroscopic/kinetic ordering at the interpolation points. In contrast, the decoder-aware curved FM has lower field error throughout the reported two-parameter cylinder test. This difference is consistent with the different conditional geometries: the cavity generator transports from a simple reference along a one-dimensional condition coordinate, while cylinder FM exploits local source cases and a decoder-metric chart in the sparse plane. Accordingly, the comparison concerns complete frozen pipelines, not an architecture-only contest between FM and diffusion. This dependence on the reference law, probability path, and conditioning construction is consistent with the distinct transport formulations underlying flow matching and diffusion[42, 43, 44, 46].
Physical adaptation also has a narrower interpretation than field reconstruction. For the cavity, access to and permits matched-grid BGK, moment, and conservation audits. The adapters reduce the registered BGK diagnostic in interpolation and upper extrapolation while approximately preserving field accuracy, but do not do so in lower extrapolation. For the cylinder, the available variables permit no-penetration, inlet, and weak-mass constraints but not a closed kinetic or momentum-energy residual. The paired frozen-versus-adapted audit therefore establishes improvements only in those registered diagnostics, whereas the out-of-sample field evaluation shows that such improvement does not guarantee a lower error at every additional condition. More specifically, the formal Physics-constrained FM model improves global boundary mass balance but not its local CV diagnostic. Its wall result comes from the analytic projection, while the learned correction contributes only as part of the jointly deployed adapter–projection system. The Physics-constrained Diffusion model improves both mass diagnostics; neither statement implies closed momentum or energy conservation. The need to balance a physics residual against fidelity to the learned data manifold is also observed in physics-informed generative modeling more broadly[54, 55, 56].
Finally, interpolation and extrapolation must be separated. At the displayed and cases, cavity errors remain near , but at the lower extrapolation point they rise to the scale for both generator families. The similar behavior of frozen and adapted predictions at this condition identifies the conditional latent generator, rather than the local field adapter, as the main bottleneck. The remedy is additional condition coverage or an extrapolation-aware transport prior, not a larger decoder correction. Generated ensemble spread is likewise retained as a numerical transport diagnostic: because each condition has one deterministic steady reference, it is not evidence of calibrated physical uncertainty. The sampling counts should be interpreted with similar care. FM and diffusion use different integration schemes, ensemble sizes, and, for the cylinder, different reference constructions, so the number of ODE or denoising steps is not by itself a fair measure of end-to-end inference cost.
Taken together, the comparisons support a component-wise interpretation of the framework. The neural field supplies a common continuous representation for heterogeneous DVM and DSMC outputs; FM and diffusion provide alternative condition-dependent transports whose relative performance changes with the parameter geometry and reference construction; and the frozen structured adapter tests whether selected physical diagnostics can be improved without retraining the representation or generator. The direct MLP comparison limits the accuracy claim: on the present single-valued steady maps, generative transport is not uniformly more accurate than deterministic regression. Its additional value is instead the explicit transport construction, sample-based inference, and compatibility with future conditional distributions containing multiple admissible or stochas-
tic states. The last capability is not validated by the present one-reference-per-condition data and should be treated as a direction for extension rather than a measured uncertainty result.
Summary and conclusions
This work presented a conditional latent generative framework for two steady rarefied-flow problems with different parameter dimensions and output structures. Neural-field auto-decoders compress the high-fidelity solutions, and train-only principal-component charts organize the optimized lookup latents. FM or diffusion then generates a condition-dependent chart endpoint directly from a simple reference distribution, without a deterministic condition-to-latent regressor as the predictive backbone. The frozen inverse chart and decoder recover the physical fields. This modular construction accommodates both structured DVM fields and native-mesh DSMC fields while separating representation, conditional-generation, and physical-correction errors.
The findings establish four distinct points. First, one neural-field representation can support structured DVM fields and native-mesh DSMC fields. Second, FM and diffusion can predict the retained latent charts, but their relative performance depends on condition-space geometry, latent metric, reference construction, and numerical transport. Third, zero-initialized structured adapters can improve selected, computable physical diagnostics while the representation and generative backbone remain frozen. Fourth, the audit answers the motivating question of where sampling-based transport adds value: for the present single-valued steady problems, deterministic regression remains a strong point predictor—five-seed MLPs at two capacity levels match or exceed the generators on several metrics (cavity interpolation means are statistically indistinguishable between the generators and lie above the MLP band; cylinder MLPs outperform both generator five-seed means over all seven additional conditions), whereas FM remains robustly above diffusion on the cylinder. The
distinctive assets of the generative formulation are therefore its sampling dimension, the frozen physics-adaptation interface, and extensibility to multivalued or stochastic solution families; the audit delineates this regime of applicability. Matching the generator capacity does not remove the deterministic advantage, so it is not a capacity artifact. The principal empirical contribution is therefore a unified study of neural-field representation, conditional latent transport, and frozen physical adaptation, not evidence that sampling-based transport is universally preferable for a single-valued steady surrogate.
The representation study showed that explained latent variance alone is not a sufficient rank-selection criterion: weak latent directions can be accurately reconstructed but remain difficult to generate from sparsely sampled physical conditions. Decoded-field screening selected a whitened rank-four chart for the cavity and a rank-12 chart for the cylinder. The retained cylinder FM model uses a decoder-metric chart, whereas its diffusion counterpart uses ordinary unwhitened PCA. Generator selection was likewise problem dependent: plain-concatenation FM was effective for the one-parameter cavity, whereas decoder-aware curved FM was the strongest frozen generator over the two-parameter cylinder domain. These are rankings of complete frozen pipelines: because the cylinder charts and reference constructions differ, they should not be interpreted as an isolated algorithmic ranking of FM against diffusion.
On the displayed cavity interpolation conditions, and , the macroscopic and reduced-distribution errors remain predominantly at the level. At the upper extrapolation condition , diffusion is more accurate than FM, whereas both generators degrade to the level at the lower extrapolation condition . The limited response of the structured adapters at the latter condition indicates that the dominant error originates in extrapolation of the latent endpoint rather than in local field decoding. For the cylinder, the equal-case six-field RMSE is 0.03655 for Frozen FM and 0.04748 for Frozen Diffusion. The corresponding contours recover the bow shock, near-wall layer, and wake, although coherent shock-aligned discrepancies and DSMC sampling texture remain visible. These results emphasize that interpolation accuracy does not by itself establish reliable continuation beyond the sampled condition support.
Physics correction was introduced through zero-initialized structured low-rank adapters while freezing the selected generator, latent transform, and principal decoder. For the cavity, normalized manifold anchoring reduces the registered matched-grid BGK diagnostic while approximately preserving macroscopic and kinetic accuracy over the displayed interpolation cases and at . The diffusion checkpoint improves its invariant diagnostics, whereas the FM checkpoint has conservation and moment ratios of 1.0574 (+5.7%); the FM result is not interpreted as improved conservation. For the cylinder, the admissible constraints are limited to wall no-penetration, inlet consistency, local control-volume mass balance, and global boundary mass balance. The formal Physics-constrained FM model enforces no-penetration exactly (wall ratio ) with essentially unchanged field error; the paired attribution shows that the wall improvement is already obtained by the analytic hard map alone, while the learned correction adds the inlet and global-balance improvements; the local control-volume ratio remains 1.0218. The learned adapter by itself is not a feasible wall model because it was trained upstream of the hard projection. The Physics-constrained Diffusion model reduces both local and global mass ratios but remains less accurate than FM in the common field audit. A lower registered physics score consequently neither guarantees a smaller field error nor establishes satisfaction of the complete Boltzmann or BGK system.
For the difficult lower-side cavity extrapolation, the principal limitation is condition-space coverage rather than spatial query resolution. Additional low-Knudsen training conditions or an extrapolation-aware latent transport prior are therefore more likely to help than increasing adapter capacity. A stronger cylinder audit would require pressure-tensor and heat-flux data to close the momentum and energy balances, together with independent DSMC realizations or averaging blocks to separate sampling variability from surrogate error. Because each present operating condition supplies only one steady reference solution, generated ensemble spread measures numerical transport sensitivity and is not calibrated physical uncertainty. Replicated stochastic data or genuinely non-unique unsteady states would be needed before that spread could be interpreted probabilistically. Overall, the results support a modular strategy in which neural fields provide the common solution representation, conditional transport supplies sampling-based cross-condition prediction, and frozen structured adapters improve selected, explicitly auditable physical diagnostics. For the present steady single-reference data, this strategy is an extensible modeling framework rather than a demonstrated replacement for the more economical deterministic MLP.
This appendix provides the audits and additional field comparisons that support the main text: endpoint-averaging noncommutativity (Section A), leave-one-condition-out chart stability (Section B), reduced-BGK numerical sensitivity (Section C), a sampling-seed audit of the frozen cavity pipelines (Section D), computational cost (Section E), and a physics-adapter seed/weight robustness audit (Section F).
Endpoint-averaging noncommutativity
The noncommutativity is quantified after generator selection using 128 endpoints, 256 fixed spatial probes, and 2,048 fixed velocity nodes at the four cavity evaluation conditions . The relative difference between and has FM macro/kinetic mean values and maxima . For diffusion the corresponding means are and the maxima are . The diffusion maximum occurs at the lower extrapolation condition , where its endpoint RMS spread is ; the three other diffusion kinetic biases are –. Latent averaging is therefore an accurate computational surrogate for decoded averaging in interpolation and upper extrapolation, but contributes a non-negligible part of the already larger low-Knudsen extrapolation error.
Leave-one-condition-out chart stability
Because this chart is estimated from only 16 lookup codes in a 24-dimensional ambient space, we additionally refit the rank-four PCA after omitting each training condition and compare its span with the full-data span. Across the 16 folds, the largest principal angle has a mean of and a maximum of ; the worst fold omits the lower endpoint . Across all 120 pairs of LOCO folds, the corresponding mean and maximum are and . Thus, the dominant four-dimensional span is reasonably stable for most omissions but is not statistically invariant at the boundary of the training support. This finite-sample chart uncertainty is separate from the decoded LOCO error in the main-text PCA ablation table and limits claims about an intrinsic four-dimensional physical manifold.
Reduced-BGK numerical sensitivity
The reductions in the main-text cavity adapter table are evaluated with the near-wall-biased probe measure used for adapter selection. To separate that measure from differentiation and spatial-quadrature effects, an independent frozen-checkpoint audit pairs each adapted prediction with its frozen baseline using the same generated latent, the same spatial points, and all 6,400 velocity nodes. Four training-support conditions, , are given equal weight. Spatial integration uses cell-centered composite midpoint quadrature on .
Across the three difference steps, the reduction spans only 0.223 percentage points for FM and 0.104 percentage points for diffusion. The sign is also unchanged for every audited condition on the , , and grids. A control is under-resolved and reverses the local ordering at the lowest conditions, so it is excluded from Table XVIII. Between and , the aggregate reduction spans 3.095 percentage points for FM and 1.698 points for diffusion. Thus, the conclusion that adaptation lowers the registered BGK diagnostic is stable to the tested difference step and adequate query-grid refinement, but the reduction magnitude is not independent of the spatial sampling measure. In particular, the near-wall-biased values 28.65% and 23.59% in the main-text cavity adapter table should not be interpreted as uniform-area residual reductions.
| Audit setting | FM reduction | Diffusion reduction |
| , grid | 16.289% | 11.274% |
| , grid | 16.286% | 11.378% |
| , grid | 16.509% | 11.366% |
| , grid | 19.382% | 13.076% |
| , grid | 17.806% | 12.011% |
Table XVIII. Numerical sensitivity of the cavity reduced-BGK reduction. Values are equal-Kn reductions relative to each paired frozen generator under uniform-area quadrature.
Sampling-seed audit of the frozen cavity pipelines
The primary checkpoints are single training fits. To separate training replication from sampling variability, we additionally repeat only the frozen cavity inference with 20 sampling seeds and 256 endpoints per seed. Over the two displayed interpolation cases, the kinetic-error mean and population standard deviation are for FM and for diffusion; over the two extrapolation cases they are and , respectively. These small sampling-seed variations do not quantify retraining uncertainty; independent five-seed generator refits addressing that question are reported in the main text.
Computational cost
Training wall times are reported from the original run logs and therefore retain their original, nonuniform hardware context. The cavity representation fit required 200,000 updates and 20.38 logged GPU hours (sum of per-update times). The cavity FM and diffusion generator fits required 1,018 and 128 s, respectively; their physics adapters required 5,755 and 5,920 s. The retained cylinder representation fit required 131,239 s, the retained FM generator screen 522 s, and the formal FM and diffusion adapters 1,486 and 4,779 s. Because accelerator identifiers were not stored by all historical jobs, these figures document cost but are not used for cross-model speed ranking.
For a controlled inference measurement, all frozen pipelines were rerun in FP32 on one NVIDIA RTX 5070 Ti Laptop GPU after warm-up, using five repeats. For one cavity query with 256 endpoints, 2,500 spatial points, and 6,400 velocity nodes, median latent-generation/full-decoding times were 0.071/0.444 s for FM and 0.049/0.658 s for diffusion. For one cylinder query on a output grid, the corresponding times were 0.134/0.090 s for the 21-sample local FM and 0.046/0.090 s for 40-sample diffusion. These are end-to-end neural inference components, excluding file I/O and plotting. The distributed TransportBench fields do not include the original DVM or DSMC job wall times and hardware records; consequently, a defensible solver-to-surrogate acceleration ratio cannot be inferred from the available artifacts and is not reported.
Physics-adapter seed and weight robustness (cavity FM)
The retained cavity FM adapter is a single fit. To test whether its conclusion depends on the training seed or on the manually set physics weight, the adapter is retrained with two additional seeds and with the physics total weight scaled by and . Because these runs were executed on a 12-GB accelerator, all five rows (including a default-seed control) share a reduced collocation profile (16 conditions and 256 collocation points per update); the retained full-profile run is listed for reference. Selection folds, steps, and all other weights are unchanged.
| Variant | BGK reduction | Invariant ratio | Adapted kinetic | Gates |
| Retained (full profile) | 28.65% | 1.0574 | feasible | |
| Default seed, reduced profile | 24.54% | 1.0432 | all pass | |
| Seed +1 | 32.56% | 1.0270 | all pass | |
| Seed +2 | 25.47% | 1.0689 | all pass | |
| Weight | 10.24% | 1.0140 | all pass | |
| Weight | 14.19% | 1.0454 | all pass |
Table XIX. Cavity FM adapter robustness to training seed and physics weight. Reductions and ratios are evaluated against the paired frozen generator on independent probes; “gates” denotes the feasibility criteria of the main text.
The BGK reduction remains positive and of the same magnitude for all seeds, the invariant ratios stay within 1.027–1.069, and the adapted kinetic error never exceeds the paired frozen value ( under the reduced profile) by more than 0.5%. Halving or doubling the physics weight changes the reduction magnitude but not the feasibility or the field accuracy. The adapter conclusion is therefore robust to seed and weight perturbations; the weight scan additionally shows that larger physics weights do not purchase larger reductions, consistent with the gate-based checkpoint selection of the main text.
Acknowledgments
Y. Qi would like to thank Dr. Tianyi Li (李天一), Miss Hui Jin (金慧), and Mr. Kuilong Chen (陈奎龙) for their support during the research process. This work was supported by the Science Foundation for Young Scientists of the State Key Laboratory of High Temperature Gas Dynamics (2025QN16), Chinese Academy of Sciences Project for Young Scientists in Basic Research (YSBR107), Strategic Priority Research Program of the Chinese Academy of Sciences (XDB0620403), National Natural Science Foundation of China (12302381 and 12572340), and Beijing Natural Science Foundation (L252039).
DATA AVAILABILITY
The source code required to reproduce the reported models and evaluations will be made publicly available upon publication together with a SHA-256 manifest for the ten frozen checkpoints used in the reported pipeline comparisons. The cavity and cylinder data are from TransportBench and are publicly available at https://huggingface.co/datasets/CFDML/TransportBench.
References
- [37]J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove, “DeepSDF: Learning continuous signed distance functions for shape representation,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) pp. 165–174.arxiv.org/abs/1901.05103
- [38]Y. Yin, M. Kirchmeyer, J.-Y. Franceschi, A. Rakotomamonjy, and P. Gallinari, “Continuous PDE dynamics forecasting with implicit neural representations,” in ICLR 2023 Conference (2023).arxiv.org/abs/2209.14855
- [39]Serrano, L., Le Boudec, L., Kassaï Koupaï, A., Wang, T. X., Yin, Y., Vittaut, J.-N., & Gallinari, P. (2023). Operator learning with neural fields: Tackling PDEs on general geometries. Advances in Neural Information Processing Systems, 36, 70581–70611.
- [40]G. Arvanitidis, L. K. Hansen, and S. Hauberg, “Latent space oddity: On the curvature of deep generative models,” in International Conference on Learning Representations (2018).arxiv.org/abs/1710.11379
- [41]J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in Proceedings of the 32nd International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 37 (PMLR, 2015) pp. 2256–2265.arxiv.org/abs/1503.03585
- [42]Ho, J., Jain, A. N., & Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33, 6840–6851.
- [43]Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., & Poole, B. (2021). Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations.
- [44]Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., & Le, M. (2023). Flow matching for generative modeling. ICLR 2023 Conference.
- [45]X. Liu, C. Gong, and qiang liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” in The Eleventh International Conference on Learning Representations (2023).arxiv.org/abs/2209.03003
- [46]A. Tong, K. FATRAS, N. Malkin, G. Huguet, Y. Zhang, J. Rector-Brooks, G. Wolf, and Y. Bengio, “Improving and generalizing flow-based generative models with minibatch optimal transport,” Transactions on Machine Learning Research (2024).arxiv.org/abs/2302.00482
- [47]Gao, H., Kaltenbach, S., & Koumoutsakos, P. (2025). Generative learning of the solution of parametric Partial Differential Equations using guided diffusion models and virtual observations. Computer Methods in Applied Mechanics and Engineering, 435, 117654.
- [48]Shukla, A., Diaconu, C., Bergamin, F., Perdikaris, P., Hernández-Lobato, J. M., Turner, R. E., & Mathieu, E. (2024). On conditional diffusion models for PDE simulations. Advances in Neural Information Processing Systems, 37, 23246–23300.
- [49]Parikh, M. H., Fan, X., & Wang, J.-X. (2026). Conditional flow matching for generative modelling of near-wall turbulence with quantified uncertainty. Journal of Fluid Mechanics, 1029, A32.
- [50]Ramos, D. L., Lacasa, L., Gutiérrez, F. R., Verelst, E., & Rubio, G. (2026). FluidFlow: A flow-matching generative model for fluid dynamics surrogates on unstructured meshes. abs/2604.08586
- [51]Armitage, V., Ramic, Y., & Mishra, S. (2025). Rectified flows for fast multiscale fluid flow modeling. abs/2506.03111
- [52]M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational Physics 378, 686–707 (2019).DOI
- [53]T. De Ryck, S. Mishra, and R. Molinaro, “wPINNs: Weak physics-informed neural networks for approximating entropy solutions of hyperbolic conservation laws,” SIAM Journal on Numerical Analysis 62, 811–841 (2024), doi: 10.1137/22M1522504.DOI
- [54]D. Shu, Z. Lin, and A. Barati Farimani, “A physics-informed diffusion model for high-fidelity flow field reconstruction,” Journal of Computational Physics 478, 111972 (2023).DOI
- [55]Jacobsen, C., Zhuang, Y., & Duraisamy, K. (2025). CoCoGen: Physically consistent and conditioned score-based generative models for forward and inverse problems. SIAM Journal on Scientific Computing, 47, C939–C425. https://doi.org/10.1139/24M1636071
- [56]Bastek, J.-H., Sun, W., & Kochmann, D. M. (2025). Physics-informed diffusion models. International Conference on Learning Representations, 2025, 3360–3385.
- [57]N. Houlsby, A. Giurgiu, B. Jastrzebski, Q. D. Morrone, A. Laroussilhe, A. Gesmundo, M. Attarian, and S. Gelly, “Parameter-efficient transfer learning for NLP,” in Proceedings of the 36th International Conference on Machine Learning, Vol. 97, edited by C. Kamalika and S. Ruslan (PMLR, 2019) pp. 2790–2799.arxiv.org/abs/1902.00751
- [58]Hu, E., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2022). LoRA: Low-rank adaptation of large language models.
- [59]Bhatnagar, P. L., Gross, E. P., & Krook, M. (1954). A model for collision processes in gases. i. small amplitude processes in charged and neutral one-component systems. Physical Review, 94, 511–525.
- [60]Perez, E., Strub, F., de Vries, H., Dumoulin, V., & Courville, A. C. (2018). FiLM: Visual reasoning with a general conditioning layer. AAAI Conference on Artificial Intelligence, 3942–3951.
- [61]Song, J., Meng, C., & Ermon, S. (2020). Denoising diffusion implicit models.
- [62]Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., & Zhu, J. (2025). DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research, 22, 730–751.
- [63]Loshchilov, I., & Hutter, F. (2019). Decoupled weight decay regularization. International Conference on Learning Representations.