betaFeedback

WorksPhysics and Astronomy

Geometry-native machine learning reconstruction of DSMC moment fields with support monitoring

Abstract

Direct simulation Monte Carlo (DSMC) resolves rarefied-gas dynamics without a constitutive closure, but finite-sample estimates of macroscopic moments converge at markedly different rates. We develop a non-intrusive, geometry-native machine learning reconstruction of the retained two-dimensional moment hierarchy: number density, two velocity components, translational temperature, three pressure-tensor components, and two heat-flux components. From three sampling blocks, the estimator corrects a structured prior learned from development data with a bounded term computed from the current observation, while preserving additive-moment consistency and the measured zero-frequency content. In cavity development tests, the final observation-conditioned estimator, whose prior is a trained MambaIR restoration network, reduces transverse-heat-flux error to 0.658 and 0.672 times that of a ten-block direct average at two rarefied conditions. For a hypersonic cylinder, a cylinder-centred estimator is fixed before evaluation on six new observation/reference pairs. It improves both global transverse heat flux and near-wall normal heat flux in every pair; the ratios of arithmetic-mean normalised root-mean-square errors (NRMSEs) are 0.846 and 0.793, and the Holm-adjusted one-sided exact probabilities are 0.03125.

Introduction

Direct simulation Monte Carlo (DSMC) is the standard particle method for rarefied and nonequilibrium gas dynamics [1]. Its macroscopic fields are statistical estimates, and their sampling errors differ substantially with moment order [2]. Number density and bulk velocity often converge long before stress and heat flux. Heat flux is a signed third-order central moment: large molecular contributions cancel to produce a comparatively weak mean, so a sampling budget that is adequate for density or temperature may remain unusably noisy for heat-transfer analysis. The computational question addressed in this paper is how much of a physically meaningful DSMC moment hierarchy can be recovered from a short sampling window by a machine-learning estimator and, no less important, how a deployed estimator can decide, without access to a reference solution, whether its reconstruction should be trusted at the current operating condition.

A mature body of work reduces DSMC noise by modifying the particle calculation itself. Control variates and deviational formulations simulate the departure from an equilibrium or reference state [3, 4, 5]; importance weights, relaxation-time and variable-hard-sphere formulations, and multi-effect or binary-mixture extensions broaden the accessible regimes [6, 7, 8, 9, 10, 11, 12]. Most recently, denoising multiscale particle methods embed low-noise collective information directly in the simulator particles [13, 14]. These methods can be highly effective, but they require a modified solver, a reference distribution, or online model integration. The problem considered here is complementary: a production DSMC solver has already written additive molecular moments, and the objective is to reconstruct the complete retained hierarchy non-intrusively, after sampling, from a short window.

Non-intrusive filtering of particle fields also has direct precedent in the literature: Zimoń et al. [15] systematically evaluated proper-orthogonal-decomposition and wavelet-type noise-reduction algorithms for DSMC and molecular-dynamics fields in multiscale applications, building on the classical estimation toolbox of low-rank projection, Wiener filtering, and transform-domain shrinkage [16, 17, 18, 19]. That generation of methods left open three questions that motivate the present work. Generic filters use no information beyond the current realisation, although repeatable flow structure across operating conditions and random seeds is an exploitable prior. A filter tuned to look smooth can also silently corrupt physically meaningful low-frequency content, the amplitude and spatial mean of a weak signed moment, which is precisely the content that matters for cancellation-dominated heat flux. Most important for deployment, none of these estimators reports when it should not be used: a prior-based reconstruction that is accurate at its development conditions may remain visually plausible, yet biased, after the operating condition shifts.

Machine-learning reconstruction of fluid fields from sparse or noisy data is now an established research direction [20]. Super-resolution and self-supervised denoising have been demonstrated for turbulence [21, 22, 23, 24], sparse-sensor reconstruction via sparse representation and physics-informed networks [25, 26, 27, 28, 29, 30, 31], and hard measurement constraints have proved decisive under noise: Mo and Magri Mo and Magri [32] show that enforcing the mean of the prediction at sensor locations is the robust choice for noisy data, and extend the approach to three-dimensional reconstruction [33]. The data-consistency operation used here belongs to the same family: the observed zero-frequency content is preserved exactly while a bounded transform-domain residual corrects the prior, a construction that inherits its rationale from regularised and learned inverse problems [34, 35]. Self-supervised training without clean targets follows the noise-to-noise principle [36, 37], and the structured priors are drawn from modern restoration backbones [38, 39]. Particle moment fields, however, differ from images and velocimetry planes in ways that this literature does not address: the noise is strongly heteroscedastic, the reported quantities share nonlinear centralisation operations acting on common additive accumulators, the natural representation changes with geometry, and physics-informed residuals face an identifiability barrier discussed next. Physics-informed and operator-learning methods [40, 41, 42, 43], including recent kinetic surrogates and learned collision models from our group [44, 45, 46, 47, 48], accelerate the forward kinetic problem; the present work instead treats the statistical inverse problem attached to an unmodified solver, and uses conservation identities diagnostically rather than as training losses, wary of the optimisation pathologies such residuals can introduce [49].

A smooth heat-flux field can carry the wrong amplitude or spatial mean, and an exact energy-balance residual cannot uniquely repair the error: in two dimensions, energy conservation constrains the divergence of the heat-flux vector but is insensitive to its solenoidal part, including a constant component offset. “Adding physics” through a conservation penalty therefore does not identify the difficult component; the missing information must come from the current observation. The estimator built here pairs a structured prior learned from development trajectories with a bounded residual formed from the present DSMC observation and preserves the observed zero-frequency content exactly. All central moments are formed only after additive accumulators have been combined, which prevents the systematic bias produced by centralising individual sampling blocks before averaging.

Reduced error at a tested condition does not establish unrestricted deployment. We accompany the reconstruction with a reference-independent support monitor: an empirical-Bayes observation gain is computed for nine field components in three geometric zones, and a fixed development-envelope threshold is calibrated by a leave-one-unit-out maximum-score rule on Mach-8 and Mach-10 trajectories. The construction parallels split-conformal calibration [50, 51, 52], applied to a familywise gain score rather than a pointwise nonconformity score, and, to our knowledge, has no precedent as an accept/abstain rule for reconstruction attached to a particle kinetic solver. Specificity is tested on held-out Mach-10 observations, and shift detection is tested on subsequently generated Mach-12 observations whose support decisions are locked before any paired reference is accessed. This separates two questions that are often conflated: whether a reconstruction is numerically useful, and whether it belongs to the regime in which its error reduction has been validated.

The method is developed on a rarefied lid-driven cavity and reformulated for hypersonic argon flow over a circular cylinder. The cavity stage uses the MambaIR state-space restoration model [39] as a structural prior and a discrete cosine transform to separate reliable low modes from sampling-dominated content; the cylinder stage uses native cell areas, cylinder-centred normal and tangential heat flux, and a bounded two-component spectral transfer. Nine two-dimensional fields are retained: number density, the two velocity components, translational temperature, three pressure-tensor components, and the two heat-flux components. A direct estimate constructed from BB additive sampling blocks is denoted Raw⁡(B)\operatorname{Raw}(B); the principal low-budget estimate is Raw⁡(3)\operatorname{Raw}(3) and is compared with Raw⁡(10)\operatorname{Raw}(10). The final cavity estimator reduces transverse-heat-flux error to 0.658 and 0.672 times the ten-block error at the two analysed conditions. The fixed cylinder estimator then improves global transverse heat flux and near-wall normal heat flux in all six independent Mach-10 pairs, with arithmetic-mean normalised root-mean-square error (NRMSE) ratios of 0.846 and 0.793 and Holm-adjusted one-sided exact probabilities of 0.03125. The support monitor, finally, accepts all six held-out Mach-10 observations and abstains on all four fresh Mach-12 observations before reference evaluation; a separately labelled same-condition Mach-12 analysis reconstructs all nine retained fields at approximately one-half of the ten-block error.

Relative to the intrusive variance-reduction line [4, 11, 13], to non-intrusive filter evaluations [15], and to learned flow reconstruction [20, 32], the paper makes five contributions:

  1. a non-intrusive, geometry-native machine learning reconstruction of the complete retained two-dimensional moment hierarchy from three sampling blocks, built on additive-first centralisation that is exact by construction;

  1. a data-consistent fusion rule in which a frozen structured prior is corrected by a bounded transform-domain residual while the measured zero-frequency content is preserved exactly, together with an identifiability analysis showing why energy conservation alone cannot recover a heat-flux component;

  1. a reference-independent domain-of-validity monitor with development-only calibration, validated by a held-out specificity test (six of six supported observations accepted) and a prospective condition-shift test (four of four fresh Mach-12 observations abstained before reference access);

  1. a prospective, seed-level statistical design, frozen estimators, disjoint observation/comparator/reference partitions, and exact paired sign tests, that separates development, attribution, and confirmation evidence and quantifies how much of each gain is supplied by the prior versus the current observation;

  1. boundary-specific statistical findings of independent practical value, including the measurement that the native wall-collision heat-flux tally retains an effective sample size an order of magnitude smaller than neighbouring volume fields.

The remainder of the paper defines the kinetic observables and estimator, describes the numerical and validation design, presents the cavity and cylinder results, and discusses bias, support, finite-reference uncertainty, and the distinction between cell-centred and wall-collision heat flux.

Kinetic observables and statistical formulation

Moment hierarchy and additive aggregation

Let f(x,ξ,t)f(\boldsymbol{x},\boldsymbol{\xi},t) be the one-particle velocity distribution, where x∈Ω⊂R2\boldsymbol{x} \in\Omega\subset\mathbb{R}^{2}, ξ∈R3\boldsymbol{\xi} \in\mathbb{R}^{3}, and tt denote position in the two-dimensional computational domain Ω\Omega, three-dimensional molecular velocity, and time. Thus the simulations are spatially two-dimensional with three velocity components. For a monatomic gas of molecular mass mmolm_{\mathrm{mol}}, define the bulk velocity u\boldsymbol{u}, peculiar velocity c=ξ−u\boldsymbol{c} = \boldsymbol{\xi} - \boldsymbol{u}, pressure tensor P\boldsymbol{P}, heat-flux vector q\boldsymbol{q}, and translational temperature TT by

n=∫R3f dξ,nu=∫R3ξf dξ,(1)n = \int_{\mathbb{R}^{3}} f\,\mathrm{d}\boldsymbol{\xi}, \qquad n\boldsymbol{u} = \int_{\mathbb{R}^{3}} \boldsymbol{\xi}f\,\mathrm{d}\boldsymbol{\xi}, \tag*{(1)}
P=mmol∫R3ccf dξ,q=mmol2∫R3∣c∣2cf dξ,(2)\boldsymbol{P} = m_{\mathrm{mol}}\int_{\mathbb{R}^{3}} \boldsymbol{c}\boldsymbol{c}f\,\mathrm{d}\boldsymbol{\xi}, \qquad\boldsymbol{q} = \frac{m_{\mathrm{mol}}}{2}\int_{\mathbb{R}^{3}} \lvert\boldsymbol{c}\rvert^{2}\boldsymbol{c}f\,\mathrm{d}\boldsymbol{\xi}, \tag*{(2)}
T=tr⁡(P)3nkB,(3)T = \frac{\operatorname{tr}(\boldsymbol{P})}{3nk_{\mathrm{B}}}, \tag*{(3)}

where nn is number density, kBk_{\mathrm{B}} is the Boltzmann constant, tr⁡\operatorname{tr} denotes the tensor trace, and juxtaposition of two vectors denotes their dyadic product. No Fourier-law or other continuum closure is invoked. The retained two-dimensional field vector is

m=(n,u,v,T,Pxx,Pxy,Pyy,qx,qy)T,(4)\boldsymbol{m} = (n,u,v,T,P_{xx},P_{xy},P_{yy},q_x,q_y)^{\mathsf{T}}, \tag*{(4)}

where uu and vv are the Cartesian xx- and yy-components of u\mathbf{u} and the superscript T\mathrm{T} denotes transpose.

The solver stores additive rather than central moments. For simulator particles indexed by pp, with statistical weights wpw_p, define

C0=∑pwp,Ci=∑pwpξp,i,Cij=∑pwpξp,iξp,j,(5)C_0 = \sum_p w_p,\qquad C_i = \sum_p w_p \xi_{p,i},\qquad C_{ij} = \sum_p w_p \xi_{p,i}\xi_{p,j}, \tag*{(5)}
E2=∑pwp∣ξp∣2,Fi=∑pwp∣ξp∣2ξp,i,(6)E_2 = \sum_p w_p \lvert\boldsymbol{\xi}_p\rvert^2,\qquad F_i = \sum_p w_p \lvert\boldsymbol{\xi}_p\rvert^2 \xi_{p,i}, \tag*{(6)}

where i,j∈{x,y,z}i,j \in\{x,y,z\} label Cartesian components. Here C0C_0, CiC_i, and CijC_{ij} are the weighted zeroth-, first-, and second-order raw moments, E2E_2 is the squared-speed moment, and FiF_i is the corresponding raw energy-flux moment. After summing these quantities over the selected blocks, ui=Ci/C0u_i = C_i/C_0 and the central third-order numerator is evaluated once as

Ji=Fi−uiE2−2∑jujCji+2C0ui∣u∣2.(7)J_i = F_i - u_i E_2 - 2\sum_j u_j C_{ji} + 2C_0u_i\lvert\mathbf{u}\rvert^2. \tag*{(7)}

The physical heat flux is obtained from JiJ_i using the solver’s common molecular-mass, cell-volume, and sampling-time factors. Centralising each block separately and then averaging would insert a different block velocity into the nonlinear polynomial in Equation (7); it is therefore not equivalent to additive-first aggregation. All fields in this work are formed by summing the accumulators in Equation (6) before any centralisation.

Heat flux is statistically demanding because the terms on the right of Equation (7) can be individually large while their signed sum is small. Modest absolute errors in bulk velocity or raw energy flux can therefore be a large fraction of qxq_x or qyq_y. This cancellation explains why spatial correlation and visual smoothness are insufficient diagnostics: a coherent offset or amplitude error can dominate the normalised field error even when the principal pattern is correct.

Heat-flux identifiability

The exact monatomic internal-energy balance is

32nkBDTDt+P:∇u+∇⋅q=0,(8)\frac{3}{2} n k_B \frac{D T}{D t} + \mathbf{P} : \nabla\mathbf{u} + \nabla\cdot\mathbf{q} = 0, \tag*{(8)}

where D/Dt=∂t+u⋅∇D/Dt = \partial_t + \mathbf{u}\cdot\nabla is the material derivative and :: denotes double contraction. Although Equation (8) is an exact kinetic moment identity, it supplies only one scalar constraint for the two in-plane heat-flux components. For any sufficiently smooth scalar field ψ\psi,

δq=∇⊥ψ=(∂ψ∂y,−∂ψ∂x),∇⋅δq=0.(9)\delta\mathbf{q} = \nabla^\perp\psi= \left(\frac{\partial\psi}{\partial y},-\frac{\partial\psi}{\partial x}\right),\qquad\nabla\cdot\delta\mathbf{q}=0. \tag*{(9)}

Thus q\mathbf{q} and q+δq\mathbf{q}+\delta\mathbf{q} have the same energy residual. A constant error in one component is a simple member of this null space. Conservation is retained as an audit of the reconstructed hierarchy, but it is the current observation, not residual minimisation, that must anchor the unidentifiable content.

Observations, references, and error measure

A direct field obtained from BB declared additive blocks is denoted Raw(BB); for example, Raw(3) is the average formed from three blocks. Its abstract observation model is

mBobs=mtrue+εB,E[εB]≃0,(10)m_{B}^{\mathrm{obs}} = m^{\mathrm{true}} + \varepsilon_{B}, \qquad\mathbb{E}[\varepsilon_{B}] \simeq0, \tag*{(10)}

where mtruem^{\mathrm{true}} is the infinite-sampling expectation, εB\varepsilon_{B} is the finite-sampling error, and E\mathbb{E} denotes expectation. The covariance of εB\varepsilon_{B} depends on particle population, moment order, temporal correlation, and block construction. The references used below are independent or leave-one-seed-out DSMC averages, not exact solutions. The common independent reference makes each comparison paired and does not preferentially enter either prediction, but the resulting error is measured against a stochastic reference, not against an exact Boltzmann solution.

For a scalar field aa, reference arefa^{\mathrm{ref}}, fluid-cell index cc, cell area AcA_c, and evaluation mask M\mathcal{M}, the area-weighted normalised root-mean-square error (NRMSE) is

NRMSE⁡A,M(a,aref)=[∑c∈MAc(ac−acref)2∑c∈MAc(acref)2]1/2.(11)\operatorname{NRMSE}_{A,\mathcal{M}}(a,a^{\mathrm{ref}}) = \left[\frac{\sum_{c\in\mathcal{M}} A_c(a_c-a_c^{\mathrm{ref}})^2}{\sum_{c\in\mathcal{M}} A_c(a_c^{\mathrm{ref}})^2}\right]^{1/2}. \tag*{(11)}

Equal-area cavity cells reduce Equation (11) to the usual unweighted expression. For evaluation unit jj, the estimator-to-comparator ratio for field aa is

ra,j=NRMSE⁡(ajest,ajref)NRMSE⁡(ajRaw(10),ajref).(12)r_{a,j} = \frac{\operatorname{NRMSE}(a_j^{\mathrm{est}},a_j^{\mathrm{ref}})}{\operatorname{NRMSE}(a_j^{\mathrm{Raw}(10)},a_j^{\mathrm{ref}})}. \tag*{(12)}

A value below one indicates lower error than the ten-block direct comparator against the same reference. The units of inference are independent trajectories or seeds, not spatial cells.

For the cylinder co-primary endpoints, the reported aggregate ratio is the ratio of arithmetic-mean NRMSEs,

Ra=Npair−1∑j=1NpairNRMSE⁡(ajest,ajref)Npair−1∑j=1NpairNRMSE⁡(ajRaw(10),ajref),(13)R_a = \frac{N_{\mathrm{pair}}^{-1}\sum_{j=1}^{N_{\mathrm{pair}}}\operatorname{NRMSE}(a_j^{\mathrm{est}},a_j^{\mathrm{ref}})}{N_{\mathrm{pair}}^{-1}\sum_{j=1}^{N_{\mathrm{pair}}}\operatorname{NRMSE}(a_j^{\mathrm{Raw}(10)},a_j^{\mathrm{ref}})}, \tag*{(13)}

where NpairN_{\mathrm{pair}} is the number of independent observation/reference pairs. This quantity differs from the geometric mean of the pairwise ratios ra,jr_{a,j} in Equation (12).

Geometry-native reconstruction and support monitoring

Prior-plus-observation estimator

Let gg identify a geometry and let Tg\mathcal{T}_g be a linear transform from native fields to geometry-adapted coefficients. For a three-block observation, define z3=Tgm3obsz_3=\mathcal{T}_g m_3^{\mathrm{obs}}. A geometry-specific coefficient prior zˉg\bar{z}_g and residual-transfer operator Hg\mathbf{H}_g give

z^=zˉg+Hg(z3−zˉg),m^pre=Tg−1z^.(14)\hat{z}=\bar{z}_g+\mathbf{H}_g(z_3-\bar{z}_g), \qquad\hat{m}^{\mathrm{pre}}=\mathcal{T}_g^{-1}\hat{z}. \tag*{(14)}

The superscript pre denotes the estimate before data-consistency projection. The operator may be diagonal by spectral mode or a small matrix coupling related components. Its singular values are restricted to [0,1][0,1], so the observation residual is never amplified. Hg=0H_g = 0 returns the prior, whereas Hg=IH_g = I returns the observation; II is the identity operator. The transform, prior, transfer coefficients, and masks are fixed before an evaluation observation is processed.

The transformed estimate is finally projected onto a declared data-consistency constraint. For native fluid cells, define the area-weighted mean of a vector field a\boldsymbol{a} as

⟨a⟩A=∑cAcac∑cAc.(15)\langle\boldsymbol{a} \rangle_A = \frac{\sum_c A_c \boldsymbol{a}_c}{\sum_c A_c}. \tag*{(15)}

Let q3obs=(qx,3obs,qy,3obs)T\boldsymbol{q}_3^{\mathrm{obs}} = (q_{x,3}^{\mathrm{obs}}, q_{y,3}^{\mathrm{obs}})^{\mathrm{T}} and q^pre=(q^xpre,q^ypre)T\widehat{\boldsymbol{q}}^{\mathrm{pre}} = (\widehat{q}_x^{\mathrm{pre}}, \widehat{q}_y^{\mathrm{pre}})^{\mathrm{T}} be the heat-flux subvectors of m3obs\boldsymbol{m}_3^{\mathrm{obs}} and m^pre\widehat{\boldsymbol{m}}^{\mathrm{pre}}, respectively. The cylinder estimate uses

q^=q^pre+(⟨q3obs⟩A−⟨q^pre⟩A),(16)\widehat{\boldsymbol{q}} = \widehat{\boldsymbol{q}}^{\mathrm{pre}} + \left(\langle\boldsymbol{q}_3^{\mathrm{obs}} \rangle_A - \langle\widehat{\boldsymbol{q}}^{\mathrm{pre}} \rangle_A\right), \tag*{(16)}

which exactly preserves the measured Cartesian mean. The corresponding cavity constraint assigns unit weight to the zero-frequency coefficient. These conditions do not assert that the three-block mean is exact; they prevent a historical prior from silently replacing a directly observed additive mode.

Cavity estimator

The cavity fields lie on a 100×100100 \times100 Cartesian grid. The structured prior is supplied by MambaIR, a deep image-restoration network built on the selective state-space (Mamba) sequence model [39, 53]. A state-space layer carries a hidden state hkh_k along a pixel sequence xkx_k through the discretised linear recurrence

hk=A‾hk−1+B‾xk,yk=Chk,(17)h_k = \overline{A} h_{k-1} + \overline{B} x_k,\qquad y_k = C h_k, \tag*{(17)}

where A‾=exp⁡(ΔA)\overline{A} = \exp(\Delta A) and B‾=(ΔA)−1(exp⁡(ΔA)−I)ΔB\overline{B} = (\Delta A)^{-1}(\exp(\Delta A) - I)\Delta B are zero-order-hold discretisations of the continuous system matrices AA and BB with step size Δ\Delta, and yky_k is the layer output. In the selective (Mamba) form, Δ\Delta, B‾\overline{B}, and CC are functions of the current input rather than constants, so the recurrence can retain or discard information according to content while its cost remains linear in sequence length. This combination of a global receptive field with linear cost is what makes a state-space backbone attractive for full-field restoration. MambaIR adapts the layer to two-dimensional fields: the image is unfolded into four one-dimensional sequences (row-wise and column-wise, each in both directions), the selective scan is applied to each, and the four outputs are merged; residual state-space blocks then add a local convolution branch and channel attention, restoring the neighbouring-pixel interactions and reducing the channel redundancy that a purely sequential scan loses [39].

No noise-free reference field exists for training, so the network is trained on the flow itself with the noise-to-noise principle: independent noisy realisations at the same operating condition serve as input/target pairs, and the minimiser of the resulting loss approaches the underlying mean field [36]. An ensemble of such networks is trained, and its ensemble mean is the prior qpri\boldsymbol{q}^{\mathrm{pri}} for the transverse heat flux. The prior suppresses high-frequency particle noise but can shrink uncertain amplitudes or offsets toward its development distribution, so it is not used without a current-observation correction. Let ZklpriZ_{kl}^{\mathrm{pri}} and Zkl(3)Z_{kl}^{(3)} be the two-dimensional discrete cosine transform (DCT) coefficients of the prior and three-block observation, respectively [18]; kk and ll are mode indices. Development trajectories supply repeatable-signal power Φklsig\Phi_{kl}^{\mathrm{sig}} and single-block sampling-noise power Φklnoise\Phi_{kl}^{\mathrm{noise}}. The final cavity estimate is

Gkl=ΦklsigΦklsig+Φklnoise/3,G00=1,(18)G_{kl}=\frac{\Phi_{kl}^{\mathrm{sig}}}{\Phi_{kl}^{\mathrm{sig}}+\Phi_{kl}^{\mathrm{noise}}/3},\qquad G_{00}=1, \tag*{(18)}
Z^kl=Zklpri+Gkl(Zkl(3)−Zklpri).(19)\widehat{Z}_{kl}=Z_{kl}^{\mathrm{pri}}+G_{kl}\left(Z_{kl}^{(3)}-Z_{kl}^{\mathrm{pri}}\right). \tag*{(19)}

We refer to this estimator as Mamba–Wiener: the prior coefficients ZklpriZ_{kl}^{\mathrm{pri}} come from the frozen MambaIR ensemble mean, and Gkl∈[0,1]G_{kl}\in[0,1] is the classical Wiener gain, the weight that minimises the expected squared error of a linear blend when signal and noise powers are known [17, 19]. Each mode therefore leans on the observation exactly to the degree that development data say it is trustworthy: Gkl→1G_{kl}\to1 where repeatable signal dominates, Gkl→0G_{kl}\to0 where three blocks are mostly noise. The division by three converts single-block noise power to that of the three-block average, and G00=1G_{00}=1 keeps the measured spatial mean untouched.

For the comparisons reported later (Table 2), the alternative estimators are defined once here. They serve three purposes: anchoring the sampling budgets, testing whether a generic single-realisation filter could deliver the same gain, and attributing the gain of Equation (19) between its prior and its observation. Raw(BB) is the direct BB-block average with no processing; Raw(3) and Raw(10) fix the low-budget and comparator anchors. TSVD/POD represents the generic filter class: it projects Raw(3) onto its leading rr singular modes, q^=∑i≤rσiuiviT\widehat{q}=\sum_{i\le r}\sigma_i u_i v_i^{\mathsf{T}}, with rr fixed on development data, and discards the noise-dominated tail [16]. It uses no historical information, so any margin of the final estimator over it must come from the learned prior or from the observation-conditioned gain. The remaining rows are ablations of Equation (19) itself. The Mamba prior row sets Gkl≡0G_{kl}\equiv0, giving Z^kl=Zklpri\widehat{Z}_{kl}=Z_{kl}^{\mathrm{pri}}: the frozen network output with the current observation ignored, which measures how far historical structure alone carries. Zero-frequency restoration only keeps Gkl≡0G_{kl}\equiv0 except G00=1G_{00}=1, so the prior is merely re-centred on the measured spatial mean; its gap to the full estimator isolates the value of the nonzero observed modes. The bounded data-consistent estimate is the intermediate stage of Equation (14) under the cavity zero-frequency constraint, the prior corrected by the bounded observation residual before the continuous per-mode Wiener allocation, so its gap to the final row measures that allocation. The cross-condition permutation feeds Equation (19) an observation from the wrong flow condition; if generic smoothing or the gain structure alone explained the improvement, this row would remain competitive, and its failure rules that explanation out. Because the final gain allocation was selected within the cavity analysis, all of these results are treated as method development rather than independent confirmation.

EstimatorKn=0.08Kn = 0.08, Ulid=350 m s−1U_{\mathrm{lid}} = 350\ \mathrm{m\ s}^{-1}Kn=0.10Kn = 0.10, Ulid=400 m s−1U_{\mathrm{lid}} = 400\ \mathrm{m\ s}^{-1}
Raw(3)1.65021.6560
Mamba prior1.10101.8861
TSVD/POD control0.96751.1223
Zero-frequency restoration only0.87201.1594
Bounded data-consistent estimate0.75060.9542
Mamba–Wiener estimate0.65750.6723
Cross-condition permutation3.27513.2932
Raw(10)1.00001.0000

Table 2. Cavity transverse-heat-flux accuracy. Entries are condition-mean qyq_y NRMSE ratios to the paired Raw(10) comparator; lower values indicate smaller error against the leave-one-seed-out reference. Each condition contains four independent evaluated seeds.

Cylinder-native estimator

Let (x,y)(x,y) be a native fluid-cell centre, (xc,yc)(x_c,y_c) the cylinder centre, RcR_c the radius, D=2RcD=2R_c the diameter, and

rc=(x−xc)2+(y−yc)2,cos⁡θ=x−xcrc,sin⁡θ=y−ycrc.(20)r_c=\sqrt{(x-x_c)^2+(y-y_c)^2},\qquad\cos\theta=\frac{x-x_c}{r_c},\qquad\sin\theta=\frac{y-y_c}{r_c}. \tag*{(20)}

The angle θ\theta is measured counter-clockwise from the downstream +x+x direction. Define the outward unit normal n=(cos⁡θ,sin⁡θ)T\mathbf{n}=(\cos\theta,\sin\theta)^{\mathsf{T}} and counter-clockwise unit tangent t=(−sin⁡θ,cos⁡θ)T\mathbf{t}=(-\sin\theta,\cos\theta)^{\mathsf{T}}. Cartesian heat flux is rotated into qn=q⋅nq_n=\mathbf{q}\cdot\mathbf{n} and qt=q⋅tq_t=\mathbf{q}\cdot\mathbf{t},

qn=qxcos⁡θ+qysin⁡θ,qt=−qxsin⁡θ+qycos⁡θ.(21)q_n=q_x\cos\theta+q_y\sin\theta,\qquad q_t=-q_x\sin\theta+q_y\cos\theta. \tag*{(21)}

Only fluid cells enter the transform. The components are interpolated to a 128×96128\times96 grid in (rc/D,θ)(r_c/D,\theta) and represented by a two-dimensional DCT. Native areas are retained for Equations (11) and (16).

Four development trajectories define the cylinder prior zˉg\bar{\mathbf{z}}_g as the mean of their ten-block coefficient fields. For distinct trajectories ii and jj, zi(3)\mathbf{z}_i^{(3)} and zi(10)\mathbf{z}_i^{(10)} denote their three- and ten-block coefficient vectors. Define the low-budget and target peer residuals dijlow\mathbf{d}_{ij}^{\mathrm{low}} and dijtar\mathbf{d}_{ij}^{\mathrm{tar}} by

dijlow=zi(3)−zj(10),dijtar=zi(10)−zj(10).(22)\mathbf{d}_{ij}^{\mathrm{low}}=\mathbf{z}_i^{(3)}-\mathbf{z}_j^{(10)},\qquad\mathbf{d}_{ij}^{\mathrm{tar}}=\mathbf{z}_i^{(10)}-\mathbf{z}_j^{(10)}. \tag*{(22)}

The radial DCT modes k=0,…,127k=0,\ldots,127 are divided into four contiguous groups with edges (0,32,64,96,128)(0,32,64,96,128), and the angular modes l=0,…,95l=0,\ldots,95 into groups with edges (0,24,48,72,96)(0,24,48,72,96). Within each resulting bin bb of this fixed 4×44\times4 partition, a two-component transfer matrix is obtained from

H~b=arg⁡min⁡H ∑i≠j∥dij,btar−Hdij,blow∥22+λb∥H∥F2,(23)\widetilde{H}_b=\underset{H}{\arg\min}\ \sum_{i\ne j}\left\lVert\mathbf{d}_{ij,b}^{\mathrm{tar}}-H\mathbf{d}_{ij,b}^{\mathrm{low}}\right\rVert_2^2+\lambda_b\lVert H\rVert_F^2, \tag*{(23)}

where ∥⋅∥2\lVert\cdot\rVert_2 and ∥⋅∥F\lVert\cdot\rVert_F are the Euclidean and Frobenius norms. If DbD_b is the matrix whose columns are the two-component samples in dij,blow\mathbf{d}_{ij,b}^{\mathrm{low}}, the fixed half-trace scaling is λb=12tr⁡(DbDbT)\lambda_b=\frac{1}{2}\operatorname{tr}(D_bD_b^{\mathsf{T}}). If H~b=UΣVT\widetilde{H}_b=U\Sigma V^{\mathsf{T}} is its singular-value decomposition, UU and VV contain the left and right singular vectors and Σ\Sigma contains the singular values. Each diagonal entry of Σ\Sigma is clipped to [0,1][0,1] before reconstructing HbH_b. The fitted nonzero-mode singular values range from 0.0882 to 0.1630, implying strong suppression of observation-dominated fluctuations. The zero-frequency matrix is the identity, followed by the native Cartesian correction in Equation (16).

For a new cylinder observation, Equation (14) is applied binwise to (qn,qt)(q_n,q_t), followed by inverse transformation, rotation to (qx,qy)(q_x,q_y), and native-cell data consistency. A prior-only result identifies the contribution of repeatable historical structure. A phase-scrambled residual retains the same spectral amplitudes but destroys alignment with the shock, near-wall layer, and wake; it tests whether the present observation contributes meaningful spatial information.

Complete-field support score

Operating-condition support is evaluated from the current observation, not from its error against a reference. A linearly extrapolated Mach-12 prior is formed from development mean fields,

m‾12ext=2m‾10−m‾8,(24)\overline{m}_{12}^{\mathrm{ext}}=2\overline{m}_{10}-\overline{m}_{8}, \tag*{(24)}

where m‾8\overline{m}_8 and m‾10\overline{m}_{10} denote the Mach-8 and Mach-10 mean fields. The monitored components are (n,u,v,T,Pxx,Pxy,Pyy,qn,qt)(n,u,v,T,P_{xx},P_{xy},P_{yy},q_n,q_t). For component index ℓ\ell, monitor-zone index ζ\zeta, and evaluation trajectory jj, an empirical-Bayes observation gain is

Kℓζ(j)=clip⁡[0,1](1−σℓζ2Vℓζres,(j)),(25)K_{\ell\zeta}^{(j)}=\operatorname{clip}_{[0,1]}\left(1-\frac{\sigma_{\ell\zeta}^{2}}{V_{\ell\zeta}^{\mathrm{res},(j)}}\right), \tag*{(25)}

where clip⁡[0,1](x)=min⁡(1,max⁡(0,x))\operatorname{clip}_{[0,1]}(x)=\min(1,\max(0,x)). With wc=Ac/∑c∈ζAcw_c=A_c/\sum_{c\in\zeta}A_c, the current residual power is

Vℓζres,(j)=∑c∈ζwc(m3,ℓcobs,(j)−m‾12,ℓcext)2.(26)V_{\ell\zeta}^{\mathrm{res},(j)}=\sum_{c\in\zeta}w_c\left(m_{3,\ell c}^{\mathrm{obs},(j)}-\overline{m}_{12,\ell c}^{\mathrm{ext}}\right)^2. \tag*{(26)}

The quantity σℓζ2\sigma_{\ell\zeta}^{2} is the corresponding area-weighted within-trajectory sampling-noise power estimated from the four three-block development draws. The monitor masks are: near wall, rc≤0.20 mr_c\leq0.20\,\mathrm{m}; wake, rc>0.20 mr_c>0.20\,\mathrm{m} and x−xc≥0x-x_c\geq0; and outer flow, rc>0.20 mr_c>0.20\,\mathrm{m} and x−xc<0x-x_c<0. These masks define the three values of ζ\zeta.

Development data define the largest supported gain Kℓζmax⁡,devK_{\ell\zeta}^{\max,\mathrm{dev}} for each of the 9×3=279\times3=27 component–zone combinations. With 1[⋅]\mathbf{1}[\cdot] denoting the indicator function, the deployment outside-envelope fraction is

sj=127∑ℓ,ζ1[Kℓζ(j)>Kℓζmax⁡,dev].(27)s_j=\frac{1}{27}\sum_{\ell,\zeta}\mathbf{1}\left[K_{\ell\zeta}^{(j)}>K_{\ell\zeta}^{\max,\mathrm{dev}}\right]. \tag*{(27)}

For reporting the separation from the envelope, the maximum gain-to-envelope ratio is

RjK=max⁡ℓ,ζKℓζ(j)Kℓζmax⁡,dev,(28)R_j^K=\max_{\ell,\zeta}\frac{K_{\ell\zeta}^{(j)}}{K_{\ell\zeta}^{\max,\mathrm{dev}}}, \tag*{(28)}

whose denominator is positive for every monitored component–zone combination in the development data. Calibration preserves the four three-block draws available within each development trajectory. For development unit ii, draw d∈{1,…,4}d\in\{1,\ldots,4\}, let Kℓζ(i,d)K_{\ell\zeta}^{(i,d)} be the gain computed from that draw and let Kℓζmax⁡,(−i)K_{\ell\zeta}^{\max,(-i)} be the envelope constructed without unit ii. Define

s~i=14×27∑d=14∑ℓ,ζ1[Kℓζ(i,d)>Kℓζmax⁡,(−i)].(29)\widetilde{s}_i=\frac{1}{4\times27}\sum_{d=1}^{4}\sum_{\ell,\zeta}\mathbf{1}\left[K_{\ell\zeta}^{(i,d)}>K_{\ell\zeta}^{\max,(-i)}\right]. \tag*{(29)}

The maximum of the Ndev=8N_{\mathrm{dev}}=8 leave-one-unit-out values in Equation (29) is used as the conservative development-envelope threshold, s⋆=47/108=0.435185s_\star=47/108=0.435185. The same numerical threshold is applied to the single-draw deployment score in Equation (27): observation jj is accepted when sj≤s⋆s_j\leq s_\star and otherwise receives an abstention. Because the calibration and deployment scores aggregate different numbers of draws, this maximum-score rule is treated as a fixed development heuristic, not as a finite-sample miscoverage guarantee. The construction borrows the logic of split-conformal calibration [50, 51, 52]: the threshold comes from development scores alone, although the score aggregated here is familywise over component–zone gains rather than a pointwise nonconformity measure.

A separate same-condition Mach-12 analysis measures reconstructability when in-condition information is available. Two 40-block reference trajectories define the pointwise prior m‾ℓc\overline{m}_{\ell c} for component ℓ\ell and native fluid cell cc. A distinct attribution map η(c)\eta(c) has three zones: near cylinder, rc≤0.20 mr_c\leq0.20\,\mathrm{m}; outer downstream, rc>0.20 mr_c>0.20\,\mathrm{m} and x−xc≥0x-x_c\geq0; and outer upstream, rc>0.20 mr_c>0.20\,\mathrm{m} and x−xc<0x-x_c<0. For the remaining trajectories,

m^ℓc=m‾ℓc+βℓ,η(c)(mℓc(3)−m‾ℓc),0≤βℓ,η(c)≤1,(30)\widehat{m}_{\ell c}=\overline{m}_{\ell c}+\beta_{\ell,\eta(c)}\left(m_{\ell c}^{(3)}-\overline{m}_{\ell c}\right),\qquad0\leq\beta_{\ell,\eta(c)}\leq1, \tag*{(30)}

where mℓc(3)m_{\ell c}^{(3)} is the three-block observation and βℓ,η\beta_{\ell,\eta} is a component–zone shrinkage coefficient. The coefficients are fitted by alternating the two calibration trajectories as prior and target, using 20 fixed three-block draws and the complementary 37 blocks as target, and clipping the resulting least-squares slopes to [0,1][0,1]. Heat flux is fitted pointwise in (qn,qt)(q_n,q_t) and rotated back to Cartesian components. This attribution analysis is kept separate from the extrapolated support decision.

Numerical setup and validation design

Flow configurations and sampling

Figure 1 summarises the estimator and its comparison data. The three-block observation is the only current-trajectory input to reconstruction. A ten-block direct average is a paired comparator, and an independently generated trajectory or leave-one-seed-out average provides the reference. The cavity uses a nested three/ten-block comparison, whereas the cylinder uses disjoint three- and ten-block windows. Additive accumulators are combined according to Equations (6) and (7) on every branch.

Image not extracted

Figure 1. Prediction and evaluation dataflow. Additive DSMC blocks are selected as a nested three/ten-block comparison for the cavity and as disjoint three- and ten-block windows for the cylinder. The observation and a fixed geometry-native prior produce the reconstructed field vector m=(n,u,v,T,Pxx,Pxy,Pyy,qx,qy)T\mathbf{m} = (n, u, v, T, P_{xx}, P_{xy}, P_{yy}, q_x, q_y)^{\mathsf{T}}. Reconstruction and the comparator are scored only against the independent reference, which has no path to prediction.

Image not extracted

Figure 2. Computational geometries and sampling coordinates. (a) Square cavity of side LL with lid speed UlidU_{\mathrm{lid}}, Knudsen number Kn\mathrm{Kn}, and dimensionless coordinates (x/L,y/L)(x/L, y/L). The block icons distinguish observation and comparator roles; the cavity three-block set is nested within its ten-block comparator. (b) Circular cylinder of diameter DD, centred at (xc,yc)(x_c,y_c), with dimensionless radial coordinate rc/Dr_c/D and polar angle θ\theta measured counter-clockwise from downstream. The shaded annuli denote the nearest-cell evaluation band and the compression layer.

The Knudsen number is Kn=λref/Lc\mathrm{Kn} = \lambda_{\mathrm{ref}}/L_c, where λref\lambda_{\mathrm{ref}} is the reference mean free path, the equilibrium reference-state value for the cavity and the freestream value for the cylinder, and the characteristic length LcL_c is the cavity side LL or cylinder diameter DD. The cavity is a 100 ×\times 100 Cartesian DSMC calculation of monatomic argon. Four independent seeds are analysed at each of two conditions: (Kn,Ulid)=(0.08,350 m s−1)(\mathrm{Kn}, U_{\mathrm{lid}}) = (0.08, 350\ \mathrm{m\ s}^{-1}) and (0.10,400 m s−1)(0.10, 400\ \mathrm{m\ s}^{-1}). Each trajectory contains ten additive sampling blocks. Blocks 0–2 define the low-budget observation, and all ten define its direct comparator. The reference for a seed is the leave-one-seed-out mean of the ten-block fields from the other three seeds at the same condition; the evaluated seed therefore contributes no samples to its reference. Two trajectories did not meet a conservative temperature-extremum stationarity diagnostic, although their mechanical, finiteness, collision, and heat-flux-specific checks were retained. Together with in-condition gain selection, this caveat is why the cavity study is used for method development rather than confirmatory inference.

The cylinder calculations use Bird’s DS2V visual DSMC program for two-dimensional and axisymmetric flows, operated here in its two-dimensional mode, on native adaptive cells. The gas is argon at Kn=0.1\mathrm{Kn} = 0.1, freestream temperature T∞=200 KT_{\infty} = 200\ \mathrm{K}, and diffuse wall temperature Tw=500 KT_w = 500\ \mathrm{K}. The freestream Mach number is M=10M = 10 and the corresponding speed is U∞=2634.1 m s−1U_{\infty} = 2634.1\ \mathrm{m\ s}^{-1}. Four development trajectories define the geometry-native model. The independent evaluation comprises six observation/reference pairs, with all 12 seeds distinct from development. For each observation, solver-output indices 100, 108, and 116 define Raw(3); ten different indices, 101–105 and 109–113, define the paired Raw(10) comparator. The same ten indices on the partner trajectory define the independent reference. Four intervening outputs are unused guards. The evaluated window spans 11.268≤tU∞/D≤11.39011.268 \le tU_{\infty}/D \le11.390, where tU∞/DtU_{\infty}/D is convective time; conclusions are restricted to this finite-time window.

The support monitor uses four Mach-8 and four Mach-10 development trajectories, each containing 40 blocks. Held-out specificity is measured on the six Mach-10 observations just described, without their reference partners. Four additional Mach-12 observations are then processed by the fixed rule. Each contains at least 40 complete outputs; the final 14 define a three-block observation, one guard, and a disjoint ten-block comparator. Independent Mach-12 references are withheld until after classification. The monitor’s form was motivated in part by earlier Mach-12 experience, so the out-of-support evidence is restricted to its performance on these four independent observations and is not presented as the discovery of a universal rule.

A separate same-condition Mach-12 analysis uses 12 evaluation trajectories and four 40-block reference trajectories. Two references define the prior and shrinkage coefficients in Equation (30); the other two form an 80-block verification target. Temporal-correlation statistics pool all four reference trajectories, whereas accuracy is evaluated only against the 80-block target. This split quantifies reconstruction when same-condition support is explicitly supplied and is not used to tune the extrapolated support decisions.

The numerical specifications required to interpret the reported comparisons are collected in Table 1. The variable-hard-sphere (VHS) and variable-soft-sphere (VSS) models describe molecular scattering; no-time-counter (NTC) denotes the collision-pair selection algorithm.

ItemCavityCircular cylinder
Solver and gas modelCartesian DSMC; monatomic argon; VHS scattering; NTC collisionsBird’s DS2V; monatomic argon; nearest-neighbour collisions; viscosity exponent 0.74; reciprocal VSS parameter 1.0
Geometry and cellsSquare of side LL; 100 ×\times 100 equal-area cellsD=0.3048 mD = 0.3048\ \mathrm{m}; domain x∈[−0.20,0.65] mx \in[-0.20, 0.65]\ \mathrm{m}, y∈[0,0.40] my \in[0, 0.40]\ \mathrm{m}; 17,312 populated native cells
Operating conditionsKn=0.08\mathrm{Kn} = 0.08 and 0.10; Ulid=350U_{\mathrm{lid}} = 350 and 400 m s−1\mathrm{m}\,\mathrm{s}^{-1}Kn=0.1\mathrm{Kn} = 0.1; M=8,10,12M = 8, 10, 12; U∞=2107.28,2634.10,3160.92 m s−1U_{\infty} = 2107.28, 2634.10, 3160.92\ \mathrm{m}\,\mathrm{s}^{-1}; T∞=200 KT_{\infty} = 200\ \mathrm{K}; n∞=4.247×1020 m−3n_{\infty} = 4.247 \times10^{20}\ \mathrm{m}^{-3} at M=12M = 12
Molecular parametersmmol=6.63×10−26 kgm_{\mathrm{mol}} = 6.63 \times10^{-26}\ \mathrm{kg}; reference diameter 4.17×10−10 m4.17 \times10^{-10}\ \mathrm{m} at 273 K; VHS exponent 0.81mmol=6.63×10−26 kgm_{\mathrm{mol}} = 6.63 \times10^{-26}\ \mathrm{kg}; reference diameter 3.595×10−10 m3.595 \times10^{-10}\ \mathrm{m} at 1000 K
Wall and particlesFully diffuse wall; Tw=300 KT_w = 300\ \mathrm{K}; initially 32 simulator particles per cellFully diffuse wall; Tw=500 KT_w = 500\ \mathrm{K}; 1.5×1061.5 \times10^{6} initial simulator particles
Sampling300 samples per block; ten retained blocks in each analysed trajectoryAdditive volume moments and a native wall-collision energy tally at each retained output; partitions defined above

Table 1. Table 1: Physical and numerical specifications. MM is freestream Mach number and n∞n_{\infty} is freestream number density. The DS2V solver uses local adaptive time steps, so no single global time increment is reported.

Endpoints and inference

The cavity endpoint is the equal-area NRMSE of transverse heat flux qyq_y. The Mach-10 co-primary endpoints are the area-weighted NRMSE of qyq_y over all fluid cells and of normal

heat flux qnq_n over cells satisfying 0≤rc−Rc≤0.05D0 \le r_c - R_c \le0.05D. Both the selected estimate and paired ten-block comparator are evaluated against the same independent reference. The result is reported as the ratio of arithmetic-mean NRMSEs in Equation (13) and as the six individual ratios in Equation (12).

For each endpoint, six improvements among Npair=6N_{\mathrm{pair}} = 6 independent pairs have one-sided exact sign-test probability 2−6=0.0156252^{-6} = 0.015625. Holm’s step-down correction across the two co-primary endpoints gives 0.03125 when both raw probabilities are equal [54]. A geometric mean of pair ratios and a descriptive Student-tt interval on their logarithms are also reported. The exact sign test is the inferential result; native cells are not treated as replicates. With only Nseed=4N_{\mathrm{seed}} = 4 cavity seeds per condition, the minimum one-sided exact probability is 0.0625, reinforcing its developmental interpretation. For the same-condition Mach-12 analysis, each displayed point is the geometric mean of 12 trajectory-level ratios. Its 95% percentile interval is formed from 20,000 bootstrap resamples of the 12 log ratios, with trajectories resampled as complete units and the geometric mean recomputed for each resample.

Support performance is evaluated by acceptance of at least five of the six held-out Mach-10 observations and abstention on all four independent Mach-12 observations. Reference-derived error is not part of either classification. Independent-reference scoring is reported separately as a diagnostic of how conservative the support rule is.

Temporal dependence is quantified by the integrated autocorrelation time

τint=1+2∑k=1k∗ρk,Beff=Bτint,(31)\tau_{\mathrm{int}} = 1 + 2\sum_{k=1}^{k^\ast} \rho_k,\qquad B_{\mathrm{eff}} = \frac{B}{\tau_{\mathrm{int}}}, \tag*{(31)}

where ρk\rho_k is the pooled lag-kk block autocorrelation, k∗k^\ast is the last lag before the first non-positive value, and BeffB_{\mathrm{eff}} is the effective number of independent blocks among BB nominal blocks. The cell-centred quantity qnq_n is distinguished from the native wall-collision heat flux qwq_w, defined as net molecular energy delivered to the wall per unit area and time.

For the Mach-12 wall statistic, sampling-scale oscillation near the upstream stagnation point θ=π\theta= \pi is represented by the symmetry-constrained local model

q~(θ)=a+b(π−θ)2,(32)\widetilde{q}(\theta) = a + b(\pi- \theta)^2, \tag*{(32)}

where θ\theta is in radians. The coefficients aa and bb are fitted separately to qwq_w and −qn-q_n over 149∘≤θ≤180∘149^\circ\le\theta\le180^\circ by inverse-variance weighted least squares. Between 149∘149^\circ and 159∘159^\circ, the raw mean and fitted trend are joined by the smoothstep weight h(ξ)=ξ2(3−2ξ)h(\xi) = \xi^2(3 - 2\xi), where ξ=(θ−149∘)/10∘\xi= (\theta- 149^\circ)/10^\circ; the fit has full weight above 159∘159^\circ. It enforces ∂q~/∂θ=0\partial\widetilde{q}/\partial\theta= 0 at stagnation, and the raw bin values remain reported alongside the fit. The displayed 95% sampling bands are q~(θ)±1.96sq(θ)/Beff\widetilde{q}(\theta) \pm1.96s_q(\theta)/\sqrt{B_{\mathrm{eff}}}, where q~(θ)\widetilde{q}(\theta) and sq(θ)s_q(\theta) are the angular-bin block mean and standard deviation. They use the held-out 80-block pool, for which Beff=6.5198B_{\mathrm{eff}} = 6.5198 for qwq_w and 54.6327 for the angular −qn-q_n profile.

Results

Cavity reconstruction

The cavity calculations isolate the central statistical difficulty. The learned prior reproduces the organisation of qyq_y but can retain a coherent amplitude and offset error, whereas Raw(3) contains the current low-frequency content together with much larger sampling noise. Enforcing the scalar identity in Equation (8) reduces its residual but cannot determine the solenoidal part of the heat-flux error in Equation (9). The reconstruction must therefore preserve coherent structure while returning observation-supported modes to the field.

The equal-area errors in Table 2 quantify this trade-off. The final Mamba–Wiener estimator has mean NRMSE ratios of 0.6575 and 0.6723 relative to the paired Raw(10) comparator at (Kn,Ulid)=(0.08,350 m s−1)(\mathrm{Kn}, U_{\mathrm{lid}}) = (0.08, 350\ \mathrm{m\ s}^{-1}) and (0.10,400 m s−1)(0.10, 400\ \mathrm{m\ s}^{-1}), respectively. Three observation blocks therefore reduce the error by roughly one third relative to ten direct blocks at both conditions. The prior alone is unreliable at the second condition, where its ratio is 1.8861; the continuous Wiener residual corrects the broad bias without restoring the full high-frequency variance of Raw(3). Cross-condition permutation gives ratios above 3.27, which rules out generic smoothing as the explanation. Because the gain is estimated within the same two-condition analysis and only four seeds are available per condition, the cavity evidence is interpreted as method development rather than confirmatory inference.

The spatial diagnostics point in the same direction and show the mechanism behind the table. In Figure 3, the reference and reconstruction share the circulation-induced structure of qyq_y, the residual bias is smooth rather than speckled, and positive root-mean-square-error (RMSE) reduction occupies most of the active region. The representative fields in Figure 4 and Figure 5 explain why the estimators in Table 2 behave so differently. Raw(3) is dominated by fine-grained sampling texture: heat flux is a signed third-order moment whose large molecular contributions cancel, so its cell estimates carry the largest relative sampling noise in the retained hierarchy, and three blocks leave that variance essentially undiluted. The Mamba prior removes this texture almost completely, which is the expected behaviour of a noise-to-noise network: trained to map one noisy realisation to another, its loss minimiser approaches the mean of the development fields, and the ensemble average suppresses the remaining seed-to-seed variability. The same training objective is also the source of its failure at the second condition: a network fitted to the development distribution pulls the broad low-frequency amplitude of an individual condition toward that distribution, so the prior stays visually convincing while carrying the coherent amplitude and offset error quantified in the table. The bias it leaves behind is smooth by construction, and smooth content is precisely what the current observation still measures with a usable signal-to-noise ratio at low mode numbers. The independently sampled Raw(10) field remains stochastic and is a finite-budget comparator, not exact truth.

Image not extracted

Figure 3. Cavity ensemble diagnostics at both conditions. The plotted variable is transverse heat flux qyq_y in W m−2^{-2}. Columns show the independent-reference mean, reconstructed mean, signed bias, and local ΔRMSE=RMSEB=10−RMSErec\Delta\mathrm{RMSE} = \mathrm{RMSE}_{B=10} - \mathrm{RMSE}_{\mathrm{rec}}. Positive ΔRMSE\Delta\mathrm{RMSE} denotes improvement over Raw(10); four independent seeds enter each row.

Image not extracted

Figure 4. Cavity qyq_y at Kn=0.08\mathrm{Kn}=0.08 and Ulid=350 m s−1U_{\mathrm{lid}}=350\ \mathrm{m\,s}^{-1}. The upper row shows the leave-one-seed-out reference, Raw(3), Mamba prior, bounded data-consistent estimate, Mamba–Wiener estimate, and paired Raw(10). The lower row gives signed error normalised by the reference root-mean-square magnitude; colour limits are identical within each row.

Image not extracted

Figure 5. Cavity qyq_y at Kn=0.10\mathrm{Kn}=0.10 and Ulid=400 m s−1U_{\mathrm{lid}}=400\ \mathrm{m\,s^{-1}}. Panel order, normalisation, and within-row colour limits are those of Figure 4. The prior recovers the spatial pattern but underestimates broad low-frequency content; conditioning on the observation reduces this bias without reintroducing the variance of Raw(3).

The profiles in Figure 6 and Figure 7 localise the correction physically. Through the near-lid layer, where the driven shear generates the strongest transverse transport, and across the interior recirculation, the conditioned estimate follows the reference amplitude that the prior underestimates, without reintroducing the block-to-block variance of Raw(3). The same profiles indicate why the generic controls in Table 2 cannot close this gap. TSVD/POD retains the most energetic singular modes of one noisy realisation, but at three blocks those leading modes already mix sampling noise into the signal, and the truncation discards weak large-scale content that a cancellation-dominated signed moment needs. Zero-frequency restoration corrects the single number its constraint protects, the spatial mean, and leaves the spatially varying low-mode amplitude error untouched.

Image not extracted

Figure 6. Ensemble-mean qyq_y along vertical cuts at x/L=0.20x/L = 0.20, 0.500.50, and 0.800.80. The ordinate y/Ly/L is dimensionless height. The upper and lower rows correspond to (Kn,Ulid)=(0.10,400 m s−1)(\mathrm{Kn}, U_{\mathrm{lid}}) = (0.10, 400\ \mathrm{m\,s}^{-1}) and (0.08,350 m s−1)(0.08, 350\ \mathrm{m\,s}^{-1}), respectively. Four seeds enter each mean and heat flux is in physical units.

Image not extracted

Figure 7. Ensemble-mean qyq_y along horizontal cuts at y/L=0.15y/L = 0.15, 0.500.50, and 0.850.85. The abscissa x/Lx/L is dimensionless position; rows and method definitions are identical to Figure 6. Observation conditioning preserves signed lateral redistribution while suppressing Raw(3) variance.

The DCT gain in Figure 8 makes the allocation explicit. Development data assign Gkl→1G_{kl} \to1 only in a compact low-(k,l)(k,l) region where repeatable signal power exceeds three-block noise power, and Gkl→0G_{kl} \to0 elsewhere, so the estimator is a per-mode bias–variance compromise: prior-based variance suppression at high frequencies and observation-anchored amplitude at low frequencies. The map also explains the severity of the cross-condition permutation: a permuted observation injects the wrong condition’s content exactly in the trusted low-frequency band, which is why its error exceeds even Raw(3) and why the gain structure alone cannot account for the improvement.

Image not extracted

Figure 8. Spectral allocation and cavity error. The first panel shows the continuous Wiener gain in the two-dimensional discrete cosine transform (DCT) basis; zero rejects and one retains an observed mode. Remaining panels show signed ensemble bias and local ΔRMSE\Delta\mathrm{RMSE} for the two conditions. Coordinates are normalised by LL, and positive ΔRMSE\Delta\mathrm{RMSE} denotes improvement over Raw(10).

Mach-10 cylinder reconstruction

Applying the cavity operator directly to the cylinder increased global qyq_y and near-wall qnq_n errors to 2.245 and 3.377 times those of Raw(10). A Cartesian cavity basis does not represent a curved wall, bow shock, compression layer, and wake in their natural coordinates. The cylinder estimator therefore retains the data-consistency principle but replaces the prior, coordinates, quadrature weights, and spectral transfer. Figure 9 shows the native areas, cylinder-centred angle, and near-wall selection. The area panel is the quadrature argument: the adaptive DS2V cells differ substantially in area between the refined near-body region and the coarse outer field, so equal-weight fitting would let the refined region dominate every spectral coefficient and misrepresent the wake. The angle panel is the alignment argument: in (rc/D,θ)(r_c/D,\theta) the bow shock, compression layer, and wall layer follow coordinate lines and a two-dimensional DCT represents them with few modes, whereas the same structures cut obliquely across a Cartesian basis and spread over many weakly determined coefficients. No solid cell enters the reconstruction, and the normal/tangential rotation in Equation (21) precedes spectral shrinkage.

Image not extracted

(a)

Image not extracted

(b)

Image not extracted

(c)

Figure 9. Figure 9: Cylinder-native representation. (a) Native DS2V fluid-cell centres and areas. (b) Polar coordinates relative to the cylinder centre, with dimensionless position and surface angle θ\theta. (c) Near-wall band within 0.05D0.05D of the surface, where DD is cylinder diameter. Solid cells are excluded and interpolated polar fields use a two-dimensional DCT basis.

In the cylinder figures, qi∗q_i^*, i∈{x,y,n,t}i \in\{x,y,n,t\}, denotes heat flux expressed in one common fixed set of scaled plotting units. The asterisk is a display convention rather than a physical nondimensionalisation; the same scale is used for every panel and method and therefore cancels from all NRMSE ratios. The representative field in Figure 10 contains the bow-shock, shoulder, and wake structure. The prior supplies most of the coherent pattern, while the bounded current-trajectory correction is small; this division of labour is what the ridge fit in Equation (23) learned, because three-block peer residuals are mostly noise, the half-trace regularisation shrinks the fitted singular values well below unity, and the operator therefore transfers only the fraction of the observation that development data could predict. Read as machine learning, the transfer matrices are deliberately underfitted: a bounded linear map estimated from four trajectories cannot chase a single realisation, which costs little where the prior is accurate and protects the estimate where it is not. The six-pair mean in Figure 11 shows improvement through both the compression layer and the wake, the two regions where refinement and signed transport differ most. Near-wall profiles and signed errors in Figure 12 and Figure 13 show reduced angular fluctuation in every evaluated pair rather than in a selected seed or sector: the gain appears as variance reduction about the reference profile rather than a change of its shape, consistent with shrinkage that acts on observation-dominated modes.

Image not extracted

Figure 10. Cylinder transverse heat flux for pair 01. The upper row shows the independent reference, Raw(3), frozen prior, selected three-block estimate, phase-scrambled control, and paired Raw(10). The lower row is signed error normalised by the reference root-mean-square magnitude. Coordinates are centred on the cylinder and normalised by DD; solid cells are masked.

Six-pair cylinder diagnostics showing reference mean, proposed mean, signed bias, and RMSE reduction

Figure 11. Six-pair cylinder diagnostics for qy∗q_y^*. Panels show independent-reference mean, selected three-block mean, signed bias, and local ΔRMSE=RMSEB=10−RMSErec\Delta\mathrm{RMSE} = \mathrm{RMSE}_{B=10} - \mathrm{RMSE}_{\mathrm{rec}}. Positive values denote improvement; coordinates are centred on the cylinder and normalised by DD.

Near-wall heat flux for six fresh cylinder pairs

Figure 12. Near-wall qn∗(θ)=q∗⋅nq_n^{*}(\theta)=\boldsymbol{q}^{*}\cdot\boldsymbol{n} for six cylinder pairs, where q∗=(qx∗,qy∗)\boldsymbol{q}^{*}=(q_x^{*},q_y^{*}) and n\boldsymbol{n} is the outward normal. Panels compare the independent reference, Raw(3), frozen prior, selected estimate, and Raw(10). This cell-centred moment differs from the wall-collision tally qwq_w.

Six-panel signed near-wall error plots for Fresh pair 01 through Fresh pair 06, showing Raw DSMC, Frozen prior, Proposed estimator, and Raw DSMC methods

Figure 13. Signed near-wall error 100(qn−qn,ref)/RMS⁡(qn,ref)100(q_n-q_{n,\mathrm{ref}})/\operatorname{RMS}(q_{n,\mathrm{ref}}) for the six cylinder pairs. Here qn,refq_{n,\mathrm{ref}} is the independent-reference profile and RMS denotes its root-mean-square magnitude over θ\theta. Common method definitions and limits permit direct pairwise comparison.

The selected estimate has a six-pair arithmetic-mean NRMSE of 0.18483 for global qyq_y and 0.04831 for near-wall qnq_n (Table 3). The aggregate ratios Ra\mathcal{R}_a in Equation (13) are 0.84642 and 0.79336, and all six pairwise ratios are below one for both endpoints. The geometric means of those six pairwise ratios are 0.84650 and 0.79558, with descriptive log-scale 95% intervals [0.83224, 0.86101] and [0.72427, 0.87391]. The exact one-sided sign-test probability is 2−6=0.0156252^{-6} = 0.015625 for each endpoint; Holm adjustment across the two tests gives 0.03125.

Panel A: six-pair mean
Estimatorqyq_y NRMSEqyq_y ratioqnq_n NRMSEqnq_n ratio
Raw(3)0.317881.455690.087611.43889
Raw(10)0.218371.000000.060891.00000
Frozen prior0.185130.847800.048360.79428
Phase control0.187990.860880.049300.80968
Selected estimate0.184830.846420.048310.79336
Panel B: ratio by independent pair
Endpoint010203040506
Global qyq_y0.846040.853150.824040.840480.864510.85136
Near-wall qnq_n0.755680.899330.717820.855370.739010.82228
Panel C: pair-level summary and inference
EndpointImprovedGeometric meanDescriptive 95% intervalHolm pp
Global qyq_y6/60.84650[0.83224, 0.86101]0.03125
Near-wall qnq_n6/60.79558[0.72427, 0.87391]0.03125

Table 3. Mach-10 cylinder accuracy against independent references. Panel A reports means over six pairs, Panel B reports selected-to-Raw(10) ratios for each pair, and Panel C gives pair-level inference; native cells are not replicates.

Most of the improvement comes from the frozen prior: relative to it, observation conditioning reduces mean NRMSE by only approximately 0.16% globally and 0.12% near the wall. The correctly phased correction nevertheless outperforms the phase control at both endpoints and preserves the observed mean exactly. The phase comparison is the informative one for attribution: both corrections carry identical spectral amplitudes, so the margin of the aligned correction in Figure 14 shows that the observation’s value lies in the spatial phase of its low modes, their registration with the shock, wall layer, and wake, rather than in amplitude content a permutation could mimic. The correction’s small magnitude agrees with the fitted singular values of 0.0882–0.1630. Downstream cuts in Figure 15, the vector magnitude in Figure 16, and the polar field in Figure 17, where qr∗=qx∗cos⁡θ+qy∗sin⁡θq_r^* = q_x^* \cos\theta+ q_y^* \sin\theta coincides with qn∗q_n^* for a circle, show that the result is not confined to one Cartesian component: the estimator acts on (qn,qt)(q_n, q_t) before rotating back, so the gain is shared by qxq_x, qyq_y, their magnitude, and the polar profile, a property a fixed Cartesian-component estimator would not have near a curved wall.

Six-panel heat-flux maps showing prior, proposed, and phase-control corrections for cylinder pair 01

Figure 14. Prior and current-observation correction for cylinder pair 01. The upper row shows prior qy∗q_y^*, selected-minus-prior Δqy∗\Delta q_y^*, and the phase-control correction. The lower row shows the corresponding near-wall qn∗q_n^* quantities. Separate matched scales distinguish the repeatable prior from the bounded correction.

Three-panel plot of transverse heat flux. Legend: Independent-reference 10–90% range; Independent reference; Raw DSMC, $B=3$; Frozen prior; Proposed estimator, $B=3$; Raw DSMC, $B=10$. Panels: (a) Wake cut $(x-x_c)/D=0.75$, (b) Wake cut $(x-x_c)/D=1.00$, (c) Wake cut $(x-x_c)/D=1.50$. Horizontal axis: Transverse heat flux, $q_y^*$. Vertical axis: Dimensionless height, $(y-y_c)/D$.

Figure 15. Transverse heat flux qy∗q_y^* along downstream cuts at (x−xc)/D=0.75(x-x_c)/D=0.75, 1.00, and 1.50. The ordinate (y−yc)/D(y-y_c)/D is dimensionless height. The shaded band spans the 10th–90th percentiles across six independent reference trajectories; method definitions are common to all panels.

Heat-flux magnitude and signed error for pair 01

Figure 16. Heat-flux magnitude ∣q∗∣=[(qx∗)2+(qy∗)2]1/2|\mathbf{q}^{*}| = [(q_{x}^{*})^{2} + (q_{y}^{*})^{2}]^{1/2} and signed error for pair 01 on the native grid. The upper row compares the independent reference, Raw(3), selected estimate, and Raw(10); the lower row is normalised by the reference root-mean-square magnitude.

Normal heat flux in native polar coordinates for the independent reference, Raw DSMC, and proposed estimator

Figure 17. Normal heat flux in native polar coordinates. The radial component qr∗=qx∗cos⁡θ+qy∗sin⁡θq_r^{*} = q_x^{*}\cos\theta+ q_y^{*}\sin\theta equals qn∗q_n^{*} for the outward normal. Panels compare the independent reference, Raw(3), and selected estimate; radius is normalised by DD and θ=180∘\theta= 180^\circ is upstream.

Support monitoring and the Mach-12 hierarchy

Development scores from Ndev=8N_{\mathrm{dev}} = 8 Mach-8 and Mach-10 trajectories give the support threshold s⋆=0.435185s_{\star} = 0.435185. All six held-out Mach-10 observations have sj=0s_j = 0 and are accepted, with RjKR_j^K from Equation (28) between 0.45 and 0.77. All four Mach-12 observations have sj=1s_j = 1: each of their 27 monitored component–zone gains is outside the development envelope and RjKR_j^K lies between 8.14 and 8.31 (Figure 18). References are absent from both decisions. Independent scoring afterwards shows that abstention is not numerical failure: relative to Raw(3), the outside-support estimate reduces NRMSE by 5.8% for the nine-field geometric mean, 15.9% for near-wall qnq_n, and 13.2% for global qyq_y. The monitor withholds an in-support interpretation rather than discarding the estimate.

Reference-independent support validation showing outside-support fractions for development leave-one-unit-out, held-out Mach 10, and fresh Mach 12 observations

Figure 18. Reference-independent support validation. Development-only calibration fixes s⋆=0.435185s_{\star} = 0.435185. The monitor accepts all six held-out Mach-10 observations at sj=0s_j = 0; their maximum gain-to-envelope ratios RjKR_j^K are 0.45–0.77. All four independently generated Mach-12 observations have sj=1s_j = 1 and are classified outside support before reference scoring.

With separate same-condition Mach-12 references, the prior and corrected estimate reduce error for every retained field and all 12 evaluation trajectories. Ratios in Table 4 range from 0.461 to 0.501, with nine-field geometric means of 0.479 and 0.481. The current-observation correction is marginal: the prior is slightly better for most fields, whereas the correction is slightly better for PxyP_{xy}, qxq_x, and near-wall qnq_n. The supported result is therefore an approximately twofold error reduction from same-condition information, not systematic superiority of a small residual correction.

FieldSame-condition priorCorrected Raw(3) estimate
Number density nn0.496060.50076
Streamwise velocity uu0.473860.47934
Transverse velocity vv0.469910.47384
Translational temperature TT0.491420.49354
Streamwise normal stress PxxP_{xx}0.485730.48591
Shear stress PxyP_{xy}0.472460.47185
Transverse normal stress PyyP_{yy}0.478600.48193
Streamwise heat flux qxq_x0.482960.48133
Transverse heat flux qyq_y0.461270.46344
Geometric mean, nine fields0.479030.48121
Near-wall normal heat flux qnq_n0.462500.46163

Table 4. Mach-12 geometric-mean NRMSE ratios to Raw(10) across 12 independent evaluation trajectories with same-condition information. The prior contains no current three-block observation; the corrected estimate adds the bounded residual in Equation (30). Lower values indicate smaller error against the independent 80-block target.

Temporal dependence in the reference pool is weak for volume moments but strong for the surface tally. Across the four 40-block Mach-12 reference trajectories, the pooled fields in Table 5 retain Beff=134B_{\mathrm{eff}} = 134–160160 effective blocks among 160 nominal blocks, whereas the native wall-collision tally has Beff=12.8B_{\mathrm{eff}} = 12.8. Pooled volume-moment autocorrelations fall to zero within one to two block lags, whereas the wall tally decorrelates slowly. The paired ratios are reported as directly observed; no model-based correction for reference noise is applied.

FieldBeffB_{\mathrm{eff}} of 160
Number density nn137.0
Streamwise velocity uu133.7
Transverse velocity vv148.3
Translational temperature TT138.4
Streamwise normal stress PxxP_{xx}148.3
Shear stress PxyP_{xy}157.3
Transverse normal stress PyyP_{yy}157.6
Streamwise heat flux qxq_x153.0
Transverse heat flux qyq_y160.0
Near-wall normal heat flux qnq_n160.0
Wall-collision heat flux qwq_w12.8

Table 5. Effective number of independent sampling blocks Beff=B/τintB_{\mathrm{eff}} = B/\tau_{\mathrm{int}} from Equation (31), pooled across the four 40-block Mach-12 reference trajectories with B=160B = 160 nominal blocks. Volume moments and near-wall qnq_n are cell-centred fields; the wall-collision tally qwq_w is the surface accumulator.

The wall tally qwq_w differs from nearest-cell qnq_n. For either angular profile, the upper-surface mean is ⟨q⟩θ=π−1∫0πq(θ) dθ\langle q\rangle_\theta= \pi^{-1}\int_0^\pi q(\theta)\,\mathrm{d}\theta, with θ\theta in radians. The symmetry-constrained treatment in Equation (32) changes the upper-surface mean by only −0.009%. The resulting means are ⟨qw⟩θ=21.6 kW m−2\langle q_w\rangle_\theta= 21.6\ \mathrm{kW}\ \mathrm{m}^{-2} and ⟨−qn⟩θ=17.3 kW m−2\langle-q_n\rangle_\theta= 17.3\ \mathrm{kW}\ \mathrm{m}^{-2} (Figure 19).

Mach-12 upper-surface heat-transfer distribution with coordinate convention, surface heat-transfer distribution, and stagnation-region audit; legend shows raw wall-collision bins $q_w$, raw nearest-cell bins $-q_n$, symmetry-constrained trend $\tilde{q}_w$, and symmetry-constrained trend $-\tilde{q}_n$

Figure 19. Mach-12 upper-surface heat transfer. (a) Conventions: nn is outward, θ=180∘\theta= 180^\circ is upstream, −qn>0-q_n > 0 denotes cell-centred transport toward the wall, and qw>0q_w > 0 denotes energy delivered to the wall. (b) Raw bins, symmetry-constrained trends, and 95% sampling bands qˉ(θ)±1.96sq(θ)/Beff\bar{q}(\theta) \pm1.96s_q(\theta)/\sqrt{B_{\mathrm{eff}}} from the held-out 80-block pool; Beff=6.5198B_{\mathrm{eff}} = 6.5198 for qwq_w and 54.632754.6327 for the angular −qn-q_n profile. (c) Stagnation enlargement with Equation (32); raw values remain visible.

Discussion

Geometry, data support, and the origin of reconstruction gains

The results support geometry-native regularisation rather than a geometry-independent image filter. In the cavity, Cartesian modes provide a natural separation between repeatable low-frequency structure and sampling-dominated content. Around the cylinder, the same physical idea is effective only after the heat-flux vector is expressed in normal and tangential components, solid cells are excluded, and native cell areas are retained. The distinction matters near a curved wall and through the bow shock, where a fixed Cartesian component mixes radial and angular structure. The transferable element of the method is therefore the prior-plus-observation construction in Equation (14), together with bounded residual gains and exact treatment of observed additive modes. The prior, coordinates, masks, and transfer operator remain specific to the geometry on which they were constructed.

The Mach-10 comparison also clarifies where the accuracy originates. The prior-only and observation-conditioned ratios are nearly identical, so most of the reduction relative to the ten-block direct comparator comes from repeatable structure learned from the development trajectories. The three-block correction is small, although it is not immaterial: it restores the measured Cartesian mean exactly, and a phase-scrambled correction is consistently worse than the spatially aligned correction. These controls show that the present observation carries usable information; they do not imply that it creates the dominant flow structure. They also expose a limitation. Because the fitted nonzero-mode gains are small, a short observation cannot be expected to repair an arbitrarily biased prior after a substantial operating-condition change.

The support score addresses that limitation by separating reconstruction from domain compatibility. All six held-out Mach-10 observations are consistent with the Mach-8/Mach-10 gain envelope, whereas all four Mach-12 observations lie outside it in every monitored field–zone combination. This separation is stronger than a marginal threshold crossing, but it should not be interpreted as an error estimator. Independent scoring shows that the abstained estimate still improves all three reported Mach-12 summaries relative to the raw three-block observation. Abstention therefore withholds the in-support accuracy claim; it does not declare the reconstructed field useless. Acceptance, in turn, indicates similarity to the calibrated gain regime and guarantees no prescribed error for an individual trajectory.

The present support result is deliberately narrow. It demonstrates supported-condition retention at Mach 10 and detection of the tested shift to Mach 12 for the same gas, Knudsen number, wall model, and cylinder geometry. Eight Mach-8/Mach-10 development trajectories are insufficient to establish a universal boundary in Mach number, and no conclusion follows automatically for a different Knudsen number, wall temperature, gas species, body shape, or transient regime. Extension to any of those settings requires new development coverage and an independent evaluation of both supported observations and relevant shifts. The separate same-condition Mach-12 analysis illustrates the complementary case in which in-condition information is explicitly available: all nine retained fields attain ratios near one-half, although the prior itself supplies nearly all of that gain. This reinforces the central role of data support without undermining the case for observation consistency.

Statistical interpretation and limitations

The cavity results identify a useful mechanism but remain developmental. The spectral gain allocation was selected within the cavity analysis, and only four seeds are available at each condition; even improvement for every seed cannot yield a one-sided exact probability below 0.0625. In addition, two trajectories did not satisfy the conservative temperature-extremum stationarity diagnostic noted in § 4. These two failures do not by themselves show that the reported qyq_y comparisons are spurious, but they prevent the cavity study from serving as an unqualified independent confirmation. Its strongest conclusions are that a vision prior can retain coherent structure while miscalibrating amplitude and offset, and that observation-supported low modes can correct this failure more effectively than smoothness or a scalar energy residual alone.

The Mach-10 evidence is stronger because it uses six disjoint observation/reference pairs and reports every pair. Both co-primary endpoints improve in all six pairs, giving Holm-adjusted one-sided sign-test probabilities of 0.03125. Nevertheless, the sample remains small, and the experiment supports only the declared finite-time state. The evaluated outputs span 11.268≤tU∞/D≤11.39011.268 \le tU_{\infty}/D \le11.390; they do not demonstrate asymptotic stationarity or behaviour at a later target such as tU∞/D=30tU_{\infty}/D = 30. Treating native cells as independent replicates would not remove this limitation because all cells in a field share the same stochastic trajectory, prior, and reconstruction operator. Additional independent pairs and later-time sampling are needed to quantify how stable the effect size is beyond the present window.

All reported errors are measured against finite stochastic references rather than an exact Boltzmann solution. Independent references prevent their particle noise from entering prediction, and using the same reference for an estimator and its comparator makes each ratio a fair paired comparison. Reference variance nevertheless contributes to both observed errors and can attenuate their separation. The cavity leave-one-seed-out reference and the independent cylinder references also have different variance structures, so their absolute NRMSE values should not be compared as if they had a common truth budget. The directly observed paired ratios remain the reproducible quantities; any correction for reference noise requires additional assumptions and is appropriately secondary.

The use of three rather than ten sampling blocks is a 70% reduction in the declared averaging-block budget. It is not a demonstrated 70% reduction in end-to-end wall-clock cost. Particle motion, collision sampling, transient evolution, input/output, and scheduler overhead do not scale linearly with the retained averaging window, while prior construction has an offline cost that must be amortised over later evaluations. A wall-clock speedup claim would require matched runs with identical initialisation and stopping criteria, together with measured simulation and reconstruction times. The present computational claim is therefore restricted to the sampling budget and the accompanying error comparison.

Finally, cell-centred normal heat flux and wall heat transfer must not be conflated. The quantity qn=q⋅nq_n = \mathbf{q} \cdot\mathbf{n} is a full-range velocity moment evaluated at the nearest fluid-cell centres, whereas qwq_w is a half-range molecular energy-exchange statistic accumulated at wall collisions. Their different locations and estimators explain why −qn-q_n need not equal qwq_w. Temporal dependence sharpens the contrast: the four-trajectory Mach-12 reference pool contains 160 nominal blocks, of which the volume moments retain 134–160 effectively independent blocks but the wall tally retains only 12.8. Surface heating therefore needs its own collision tally, its own autocorrelation analysis, and its own sign convention and uncertainty estimate. The symmetry-constrained stagnation treatment changes the surface mean by less than 0.01% and leaves the raw angular bins visible, but it does not convert the nearest-cell moment into a wall boundary measurement.

Conclusions

This work introduced a non-intrusive, geometry-native machine learning reconstruction framework for the complete retained DSMC moment hierarchy. The estimator combines structure learned from independent trajectories with a bounded residual from the current low-budget observation, while preserving directly observed zero-frequency content. Forming central moments only after additive accumulators are combined maintains consistency among density, velocity, temperature, pressure, and heat flux. The formulation also makes explicit why energy conservation alone cannot identify an individual heat-flux component: divergence-free errors are invisible to the scalar energy balance.

The cavity study showed that machine-learned image restoration alone can recover spatial structure yet carry appreciable amplitude and offset bias. With the observation-conditioned spectral correction added, the transverse-heat-flux NRMSE fell to 0.658 and 0.672 times the ten-block value at the two conditions considered. For the circular cylinder, reformulating the estimator in normal and tangential coordinates produced a stronger independent test: the three-block estimate improved both global transverse heat flux and near-wall normal heat flux in all six Mach-10 observation/reference pairs. The arithmetic-mean NRMSE ratios were 0.846 and 0.793, and the Holm-adjusted one-sided exact probability was 0.03125 for each endpoint.

The support monitor complemented reconstruction accuracy with a deployment decision that did not use a reference. It accepted all six held-out Mach-10 observations and abstained on all four fresh Mach-12 observations before their references were accessed, demonstrating separation between supported and shifted operating conditions under the fixed rule. Post-decision scoring nevertheless showed that an abstained estimate can remain numerically useful, whereas the separate same-condition Mach-12 analysis reduced the error of all nine retained fields to roughly half the ten-block value. Support and accuracy are therefore related but not interchangeable, and they should be reported separately.

The demonstrated reduction from ten to three sampling blocks concerns the averaging budget, not total solver wall time, and the conclusions remain restricted to the geometries, conditions, and finite-time windows examined here. Extension to other configurations requires geometry-appropriate coordinates, newly calibrated condition-specific statistics, and independent evaluation. Surface heating also requires its native wall-collision tally and temporal-correlation analysis rather than substitution by the nearest cell-centred moment. Within these limits, the results show that a single computational framework can reconstruct the complete retained DSMC field vector from a reduced sampling budget, keep its moments kinetically consistent, withstand independent validation, and report, from the observation alone, whether an estimate lies inside its validated domain.

Data and code availability

Analysis code and compact data required to reproduce the figures and metrics are available at https://github.com/Ehsan-Roohi/DSMC_Python; raw DSMC checkpoints are available from the author on reasonable request.

References

  1. [1]Graeme A. Bird. Molecular Gas Dynamics and the Direct Simulation of Gas Flows. Clarendon Press, Oxford, 1994. ISBN 9780198561958.
  2. [2]Nicolas G. Hadjiconstantinou, Alejandro L. Garcia, Martin Z. Bazant, and Gang He. Statistical error in particle simulations of hydrodynamic phenomena. Journal of Computational Physics, 187(1):274–297, 2003. doi:10.1016/S0021-9991(03)00099-8. URL https://doi.org/10.1016/S0021-9991(03)00099-8.
  3. [3]Lowell L. Baker and Nicolas G. Hadjiconstantinou. Variance reduction for Monte Carlo solutions of the Boltzmann equation. Physics of Fluids, 17:051703, 2005. doi:10.1063/1.1899210. URL https://doi.org/10.1063/1.1899210.
  4. [4]Thomas M. M. Homolle and Nicolas G. Hadjiconstantinou. A low-variance deviational simulation Monte Carlo for the Boltzmann equation. Journal of Computational Physics, 226(2):2341–2358, 2007. doi:10.1016/j.jcp.2007.07.006. URL https://doi.org/10.1016/j.jcp.2007.07.006.
  5. [5]Thomas M. M. Homolle and Nicolas G. Hadjiconstantinou. Low-variance deviational simulation Monte Carlo. Physics of Fluids, 19:041701, 2007. doi:10.1063/1.2717721. URL https://doi.org/10.1063/1.2717721.
  6. [6]Gregg A. Radtke and Nicolas G. Hadjiconstantinou. Variance-reduced particle simulation of the Boltzmann transport equation in the relaxation-time approximation. Physical Review E, 79:056711, 2009. doi:10.1103/PhysRevE.79.056711. URL https://doi.org/10.1103/PhysRevE.79.056711.
  7. [7]Hamza A. Al-Mohssen and Nicolas G. Hadjiconstantinou. Low-variance direct Monte Carlo simulations using importance weights. ESAIM: Mathematical Modelling and Numerical Analysis, 44(5):1069–1083, 2010. doi:10.1051/m2an/2010052. URL https://doi.org/10.1051/m2an/2010052.
  8. [8]Nicolas G. Hadjiconstantinou, Gregg A. Radtke, and Lowell L. Baker. On variance-reduced simulations of the Boltzmann transport equation for small-scale heat transfer applications. Journal of Heat Transfer, 132:112401, 2010. doi:10.1115/1.4002028. URL https://doi.org/10.1115/1.4002028.
  9. [9]Gregg A. Radtke, Nicolas G. Hadjiconstantinou, and Wolfgang Wagner. Low-noise Monte Carlo simulation of the variable hard sphere gas. Physics of Fluids, 23:030606, 2011. doi:10.1063/1.3558887. URL https://doi.org/10.1063/1.3558887.
  10. [10]Lajos Szalmas. Variance-reduced DSMC for binary gas flows as defined by the McCormack kinetic model. Journal of Computational Physics, 231(9):3723–3738, 2012. doi:10.1016/j.jcp.2012.01.016. URL https://doi.org/10.1016/j.jcp.2012.01.016.
  11. [11]Mohsen Sadr and Nicolas G. Hadjiconstantinou. A variance-reduced direct Monte Carlo simulation method for solving the Boltzmann equation over a wide range of rarefaction. Journal of Computational Physics, 472:111677, 2023. doi:10.1016/j.jcp.2022.111677. URL https://doi.org/10.1016/j.jcp.2022.111677.
  12. [12]Mohsen Sadr and Nicolas G. Hadjiconstantinou. Variance reduced particle solution of the Fokker–Planck equation with application to rarefied gas and plasma dynamics. Journal of Computational Physics, 492:112402, 2023. doi:10.1016/j.jcp.2023. 112402. URL https://doi.org/10.1016/j.jcp.2023.112402.
  13. [13]Hao Yang, Kaikai Feng, Ziqi Cui, and Jun Zhang. A denoising multiscale particle method for nonequilibrium flow simulations. Journal of Computational Physics, 537:114096, 2025. doi:10.1016/j.jcp.2025.114096. URL https://doi.org/10.1016/j.jcp.2025.114096.
  14. [14]Hao Yang and Jun Zhang. Multiscale simulation Monte Carlo method: Linking spatiotemporal discretization error analysis to algorithmic development. Journal of Computational Physics, 556:114799, 2026. doi:10.1016/j.jcp.2026.114799. URL https://doi.org/10.1016/j.jcp.2026.114799.
  15. [15]M. J. Zimón, R. Prosser, D. R. Emerson, M. K. Borg, D. J. Bray, L. Grinberg, and J. M. Reese. An evaluation of noise reduction algorithms for particle-based fluid simulations in multi-scale applications. Journal of Computational Physics, 325:380–394, 2016. doi:10.1016/j.jcp.2016.08.021.DOI
  16. [16]Gal Berkooz, Philip Holmes, and John L. Lumley. The proper orthogonal decomposition in the analysis of turbulent flows. Annual Review of Fluid Mechanics, 25:539–575, 1993. doi:10.1146/annurev.fl.25.010193.002543. URL https://doi.org/10.1146/annurev.fl.25.010193.002543.
  17. [17]Norbert Wiener. Extrapolation, Interpolation, and Smoothing of Stationary Time Series: With Engineering Applications. The MIT Press, Cambridge, MA, 1949. ISBN 9780262230025. URL https://mitpress.mit.edu/9780262230025/extrapolation-interpolation-and-smoothing-of-stationary-time-series/.
  18. [18]Nasir Ahmed, T. Natarajan, and K. R. Rao. Discrete cosine transform. IEEE Transactions on Computers, C-23(1):90–93, 1974. doi:10.1109/T-C.1974.223784. URL https://doi.org/10.1109/T-C.1974.223784.
  19. [19]Julianne Chung, Matthias Chung, and Dianne P. O’Leary. Designing optimal spectral filters for inverse problems. SIAM Journal on Scientific Computing, 33(6):3132–3152, 2011. doi:10.1137/100812938. URL https://doi.org/10.1137/100812938.
  20. [20]Kai Fukami, Koji Fukagata, and Kunihiko Taira. Super-resolution analysis via machine learning: a survey for fluid flows. Theoretical and Computational Fluid Dynamics, 37:421–444, 2023. doi:10.1007/s00162-023-00663-0.DOI
  21. [21]Kai Fukami, Koji Fukagata, and Kunihiko Taira. Super-resolution reconstruction of turbulent flows with machine learning. Journal of Fluid Mechanics, 870:106–120, 2019. doi:10.1017/jfm.2019.238. URL https://doi.org/10.1017/jfm.2019.238.
  22. [22]Kai Fukami, Koji Fukagata, and Kunihiko Taira. Machine-learning-based spatio-temporal super resolution reconstruction of turbulent flows. Journal of Fluid Mechanics, 909:A9, 2021. doi:10.1017/jfm.2020.948. URL https://doi.org/10.1017/jfm.2020.948.
  23. [23]H. Kim, J. Kim, S. Won, and C. Lee. Unsupervised deep learning for super-resolution reconstruction of turbulence. Journal of Fluid Mechanics, 910:A29, 2021. doi:10.1017/jfm.2020.1028. URL https://doi.org/10.1017/jfm.2020.1028.
  24. [24]Linqi Yu, Mustafa Z. Yousif, Dan Zhou, Meng Zhang, Jungsu Lee, and Hee-Chang Lim. Self-supervised learning for effective denoising of flow fields. Physics of Fluids, 36:105159, 2024. doi:10.1063/5.0231805. URL https://doi.org/10.1063/5.0231805.
  25. [25]Jared L. Callaham, Kazuki Maeda, and Steven L. Brunton. Robust flow reconstruction from limited measurements via sparse representation. Physical Review Fluids, 4:103907, 2019. doi:10.1103/PhysRevFluids.4.103907. URL https://doi.org/10.1103/PhysRevFluids.4.103907.
  26. [26]Han Gao, Luning Sun, and Jian-Xun Wang. Super-resolution and denoising of fluid flow using physics-informed convolutional neural networks without high-resolution labels. Physics of Fluids, 33:073603, 2021. doi:10.1063/5.0054312. URL https://doi.org/10.1063/5.0054312.
  27. [27]M. F. Fathi, I. Perez-Raya, A. Baghaie, P. Berg, G. Janiga, A. Arzani, and R. M. D’Souza. Super-resolution and denoising of 4D-flow MRI using physics-informed deep neural nets. Computer Methods and Programs in Biomedicine, 197:105729, 2020. doi:10.1016/j.cmpb.2020.105729. URL https://doi.org/10.1016/j.cmpb.2020.105729.
  28. [28]Dule Shu, Zijie Li, and Amir Barati Farimani. A physics-informed diffusion model for high-fidelity flow field reconstruction. Journal of Computational Physics, 478:111972, 2023. doi:10.1016/j.jcp.2023.111972. URL https://doi.org/10.1016/j.jcp.2023.111972.
  29. [29]Álvaro Moreno Soto, Alejandro Güemes, and Stefano Discetti. Complete flow characterization from snapshot PIV, fast probes and physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 419:116652, 2024. doi:10.1016/j.cma.2023.116652. URL https://doi.org/10.1016/j.cma.2023.116652.
  30. [30]Conor Rowan, Kurt Maute, and Alireza Doostan. Physics-informed solution reconstruction in elasticity and heat transfer using the explicit constraint force method. Computer Methods in Applied Mechanics and Engineering, 446(Part B):118260, 2025. doi:10.1016/j.cma.2025.118260. URL https://doi.org/10.1016/j.cma.2025.118260.
  31. [31]Pietro Cestola, Luciano Teresi, and Antonio De Simone. A flow-aware training strategy for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 455:118910, 2026. doi:10.1016/j.cma.2026.118910. URL https://doi.org/10.1016/j.cma.2026.118910.
  32. [32]Yaxin Mo and Luca Magri. Reconstructing unsteady flows from sparse, noisy measurements with a physics-constrained convolutional neural network. Physical Review Fluids, 10:034901, 2025. doi:10.1103/PhysRevFluids.10.034901. URL https://doi.org/10.1103/PhysRevFluids.10.034901.
  33. [33]Yaxin Mo and Luca Magri. Reconstruction of three-dimensional turbulent flows from sparse and noisy planar measurements: A weight-sharing neural network approach. Data-Centric Engineering, 7:e5, 2026. doi:10.1017/dce.2026.10038. URL https://doi.org/10.1017/dce.2026.10038.
  34. [34]Simon Arridge, Peter Maass, Ozan Öktem, and Carola-Bibiane Schönlieb. Solving inverse problems using data-driven models. Acta Numerica, 28:1–174, 2019. doi:10.1017/S0962492919000059. URL https://doi.org/10.1017/S0962492919000059.
  35. [35]Hemant K. Aggarwal, Merry P. Mani, and Mathews Jacob. MoDL: Model-based deep learning architecture for inverse problems. IEEE Transactions on Medical Imaging, 38(2):394–405, 2019. doi:10.1109/TMI.2018.2865356. URL https://doi.org/10.1109/TMI.2018.2865356.
  36. [36]Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2Noise: Learning image restoration without clean data. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 2965–2974. PMLR, 2018. URL https://proceedings.mlr.press/v80/lehtinen18a.html.arxiv.org/abs/1803.04189
  37. [37]Joshua Batson and Loic Royer. Noise2Self: Blind denoising by self-supervision. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 524–533. PMLR, 2019. URL https://proceedings.mlr.press/v97/batson19a.html.arxiv.org/abs/1901.11365
  38. [38]Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision – ECCV 2022, volume 13667 of Lecture Notes in Computer Science, pages 17–33. Springer, 2022. doi:10.1007/978-3-031-20071-7_2. URL https://doi.org/10.1007/978-3-031-20071-7_2.
  39. [39]Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. MambaIR: A simple baseline for image restoration with state-space model. In Computer Vision – ECCV 2024, pages 222–241. Springer, 2024. doi:10.1007/978-3-031-72649-1_13. URL https://doi.org/10.1007/978-3-031-72649-1_13.
  40. [40]Maziar Raissi, Paris Perdikaris, and George Em 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:10.1016/j.jcp.2018.10.045. URL https://doi.org/10.1016/j.jcp.2018.10.045.
  41. [41]George Em Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics, 3:422–440, 2021. doi:10.1038/s42254-021-00314-5. URL https://doi.org/10.1038/s42254-021-00314-5.
  42. [42]Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3:218–229, 2021. doi:10.1038/s42256-021-00302-5. URL https://doi.org/10.1038/s42256-021-00302-5.
  43. [43]Qin Lou, Xuhui Meng, and George Em Karniadakis. Physics-informed neural networks for solving forward and inverse flow problems via the Boltzmann–BGK formulation. Journal of Computational Physics, 447:110676, 2021. doi:10.1016/j.jcp.2021.110676. URL https://doi.org/10.1016/j.jcp.2021.110676.
  44. [44]Ehsan Roohi, Ahmad Shoja-sani, and Stefan Stefanov. Physics constrained neural collision operators for hard sphere surrogates and ab initio angle prediction in direct simulation Monte Carlo. Physics of Fluids, 38:057123, 2026. doi:10.1063/5.0328463. URL https://doi.org/10.1063/5.0328463.
  45. [45]Ehsan Roohi and Ahmad Shoja-Sani. Data-driven surrogate modeling of DSMC solutions using deep neural networks. Aerospace Science and Technology, 168(Part A):110785, 2026. doi:10.1016/j.ast.2025.110785. URL https://doi.org/10.1016/j.ast.2025.110785.
  46. [46]Ehsan Roohi, Ahmad Shoja-Sani, and Fahimeh Ebrahimzadeh Azghadi. Neural networks for rarefied gas dynamics: Relaxation problem, polyatomic shock waves, and hypersonic cylinder flow. Physics of Fluids, 38:057108, 2026. doi:10.1063/5.0334590. URL https://doi.org/10.1063/5.0334590.
  47. [47]Ehsan Roohi and Amirmehran Mahdavi. Analysis of the rarefied flow at micro-step using a DeepONet surrogate model with a physics-guided zonal loss function. Microfluidics and Nanofluidics, 30:44, 2026. doi:10.1007/s10404-026-02899-8. URL https://doi.org/10.1007/s10404-026-02899-8.
  48. [48]Ehsan Roohi. Accelerating kinetic Fokker–Planck simulations via a GPU-native deep neural network surrogate: Application to rarefied internal and hypersonic external flows. Journal of Computational Physics, 566:115261, 2026. doi:10.1016/j.jcp.2026.115261. URL https://doi.org/10.1016/j.jcp.2026.115261.
  49. [49]Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why PINNs fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022. doi:10.1016/j.jcp.2021.110768. URL https://doi.org/10.1016/j.jcp.2021.110768.
  50. [50]Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic Learning in a Random World. Springer, New York, 2005. doi:10.1007/b106715.DOI
  51. [51]Glenn Shafer and Vladimir Vovk. A tutorial on conformal prediction. Journal of Machine Learning Research, 9:371–421, 2008.arxiv.org/abs/0706.3188
  52. [52]Anastasios N. Angelopoulos and Stephen Bates. Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning, 16(4):494–591, 2023. doi:10.1561/2200000101.DOI
  53. [53]Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. doi:10.48550/arXiv.2312.00752. URL https://doi.org/10.48550/arXiv.2312.00752.
  54. [54]Sture Holm. A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics, 6(2):65–70, 1979. ISSN 03036898. URL https://www.jstor.org/stable/4615733.DOI

Paper details

vlnc:W_eyqv764q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents