betaFeedback

WorksPhysics and Astronomy

Where the Error Lives: Separating Lid-Edge and Body-Induced Error in Neural Pressure Prediction for 3D Cavity Flow

Abstract

Can a pressure surrogate score much better globally without better resolving the pressure around immersed bodies? We test this in a three-dimensional lid-driven cavity with an unregularized lid and centered pressure at native 1283128^3 resolution. On 60 cases from 12 unseen geometries, with predictions sealed before the CFD truth existed, adding a body-free, Reynolds-conditioned reference correction to the same frozen 64364^3 surrogate lifted to 1283128^3 lowers the mean full-volume relative L2L_2 error from 20.6% to 7.5%. In the post hoc regional analysis, the lid-edge junction occupies 1.5% of the fluid but holds 69% of the truth pressure energy and 95% of the uncorrected surrogate's squared error. Interior error changes by just +0.16% relative, and object-local peak error and recall change little on average. Retrospective ablations locate the learned term's contribution in the interior and near the bodies; rebasing on the reference lowers global error further with little change in cohort-mean near-body metrics. Reference-relative body-response error remains about 25% in the interior near-body band. Grid-refinement, lid-regularization and convergence checks delimit the numerical scope. In this benchmark, a large global gain chiefly reflects correction of a shared boundary field, not better resolution of the body response. We recommend reporting regional truth-energy and model-error shares alongside near-body and reference-relative body-response errors.

Introduction

Learned surrogates model flow in complex three-dimensional domains (Li et al., 2023; Rabeh et al., [37]); volumetric pressure prediction is one such task (Pajaziti et al., [35]), and FlowBench provides volumetric pressure data (Tali et al., [46]). Their accuracy is usually summarized by one relative L2L_2 error over the fluid

AI assistance. OpenAI ChatGPT/Codex and Anthropic Claude tools assisted with research planning, code development, analysis, figure preparation, drafting and revision. Reported numerical results come from executed computations, not model-generated assertions. Responsibility for the work remains with the human authors. volume, sometimes with an error near the body. Such a summary does not, by itself, establish how accurately a surrogate captures the pressure response near immersed bodies. Evaluation practice in scientific machine learning has been criticized for weak baselines and optimistic reporting [30], and aggregate metrics are known to hide failures on important subsets [33]. Here we isolate a specific source of this evaluation gap: for a transferred coarse pressure surrogate, a geometry-shared lid-edge field dominates the global error, while accuracy near the bodies follows a different pattern. Our main contribution is to locate and distinguish these errors. A reference correction serves as a controlled intervention that makes the distinction measurable, rather than as evidence that a lower global score necessarily means better body-conditioned predictions.

The setting is the lid-driven cavity with immersed obstacles, used as a benchmark for geometry-conditioned surrogates [46, 37, 29]. Where the moving lid meets a stationary wall perpendicular to its motion, the velocity boundary condition is discontinuous, and the continuum pressure is singular [47, 43, 6]. On a finite-volume grid this singularity appears as large pressures in the few cells along those lid edges. Figure 1 shows the consequence for native 1283128^3 reference fields. The lid-edge junction, 1.5% of the fluid cells, carries 69% of the centered pressure energy. A surrogate trained at 64364^3 and transferred to the native grid does not fully reproduce this structure, and 95% of its squared error sits there.

Figure 1. Where the error lives (cohort A, 60 native 1283128^3 cases). a, Centered CFD pressure magnitude relative to the case’s fluid RMS on the xyxy plane z=43.5/128z = 43.5/128 of cohort-A heterogeneous triplet-03 at Re=300\mathrm{Re} = 300 (log scale; grey, solid). The lid moves in +x+x. White boxes mark the junction cells J\mathcal{J} where this plane crosses the two lid edges perpendicular to the motion; the dashed line bounds the 8-cell wall shell. Black curves are sign changes of the centered pressure, where ∣p∣|p| falls below 1% of the RMS. b, Shares of fluid cells, truth pressure energy and the lifted surrogate’s squared error by region (case means; segments below 12% are unlabelled except the 1.5% junction cell share; values in Table 7). c, Regional relative error of three compositions. A marker’s colour names the component it contains (blue, learned; red, reference); an arrow’s colour names the component added on the way to ours (black dot). Where no arrow is drawn, the component changes that regional error by less than a marker width. All regional results are POST HOC; the no-learning control is RETROSPECTIVE. Region definitions are in § 3.3.

The reference correction lets us distinguish a boundary-driven global gain from the accuracy supplied by learning near the bodies. The chronology matters: the work began with a method, and the regional analysis followed. During development we found that the coarse-to-fine error of a 64364^3 surrogate concentrates near the walls (Section H). We therefore compose a frozen learned 64364^3 surrogate with the part of a Reynolds-conditioned empty-cavity pressure that block averaging to 64364^3 discards. The reference is fitted offline to 17 native-grid simulations, and the composition needs no retraining and no CFD at query time (§ 4). Only after the prospective test did we decompose its results by region. The paper’s principal contribution is this evaluation finding; the controlled intervention, ablations and numerical checks provide distinct layers of evidence:

  1. Principal finding: regional evaluation changes the interpretation. In this configuration, both truth energy and the transferred coarse surrogate’s squared error concentrate at the lid edges (§ 5.2). Their shares are measured separately: large truth-energy weight alone does not force a model’s error to concentrate there. The regional decomposition of Evol2E_{\mathrm{vol}}^{2} explains why global ranking can obscure body-conditioned accuracy and motivates region-resolved reporting (Equation (4)).

  1. Controlled intervention: a prospective reference comparison. Adding the body-free reference to the frozen surrogate substantially lowers global, gradient and peak-region errors on 60 cases from 12 unseen geometries (§ 5.1). The regional analysis locates that gain rather than interpreting it as new local capability. The second cohort provides a POST HOC reconstructed contrast. A retrospective rebasing baseline does better on every global metric, so the correction is not claimed to be optimal (§ 5.3).

  1. Supporting ablations: complementary roles and a local exception. Removing only the learned term identifies its substantial interior and near-body contribution (§ 5.4). One of 24 geometries across the two cohorts materially exceeds the local tolerances when the reference is added: its near-body band reaches furthest into the lid-edge junction (§ 5.5). High-Reynolds-number object-local errors remain unresolved.

  1. Supporting numerical checks: scope and limitations. Empty-cavity solutions at 64364^{3}, 1283128^{3} and 2563256^{3}, a regularized-lid cohort and more tightly converged test-truth solutions separate grid, lid-profile and iterative-convergence effects (§ 5.7). These qualify the diagnostic finding rather than establish universal portability of the correction.

Evidence status. We label every quantitative result by its relation to the test data:

  • PROSPECTIVE: predictions sealed with hashes, together with the evaluator, before the CFD truth existed;

  • POST HOC: re-analyses of such frozen predictions after the truth was opened, for example by region;

  • RETROSPECTIVE: arms or retrainings created after the truth had been inspected.

Only two results are prospective. The first is the comparison of the lifted surrogate with the reference-corrected predictor on cohort A. The second is the absolute scores of the reference-corrected predictor on cohort B. The protocol follows the logic of preregistration ([32]).

Scope. Inputs are geometry and a scalar Reynolds number. The output is kinematic pressure over the fluid volume, centered on the fluid mean, compared against a fixed OpenFOAM discretization at 1283128^{3}. We do not predict surface pressure coefficients, forces or absolute pressure. Our reference fields are numerical solutions under one discretization, not continuum truth.

Related work

Geometry-conditioned pressure fields. Volumetric pressure prediction from geometry has been studied for aortic and vascular flow ([35, 8]), parameterized vehicle and exterior flows ([19, 2] (Gu et al. 2024)), turbomachinery ([52]) and immersed swimmers ([14]). DrivAerNet++ provides volumetric flow and surface-field data, with surface-pressure benchmarks ([13]). Operator-learning architectures address boundary–interior communication and irregular domains ([56, 50, 27, 38, 54, 51, 44] (Li et al. 2023)). None of these results is directly comparable to ours: domains, boundary conditions, pressure gauges and splits differ, and we compare only with baselines trained under our contract.

Benchmarks and evaluation. FlowBench uses signed-distance conditioning for flow around objects in a lid-driven cavity and includes near-body evaluation ([46]). Geometric-DeepONet uses its three-dimensional data and global/near-body metrics, but predicts velocity only in the cited version, despite pressure being present in the dataset [37]. Related cavity pressure tasks appear in CFDBench, ArGEnT, OTNO and a journal benchmark [29, 9, 36] (Li et al. 2025). Broader benchmarks report boundary or frequency-resolved errors [45, 34], GeoABC distinguishes a region’s error reduction from its contribution to the global reduction [55], AirfRANS shows that global field errors can disagree with force-level accuracy [5], and pressure-channel relative errors can substantially exceed velocity errors under a shared training objective in airfoil flow [48]. PINNacle regularizes the cavity lid so that its velocity vanishes at the corners [18]. In a two-dimensional deep cavity with a regularized lid, Roohi & McDevitt [40] show that the ranking of PINN training workflows varies across pressure-evaluation regions. In climate modeling, Kwa et al. [22] show that a state-independent correction improves global precipitation RMSE mainly over ocean while slightly worsening it over land. We test a different intervention: with the learned surrogate held fixed, we add a geometry-shared lid-edge pressure correction tied to one solver and grid pair and measure its effect separately from the body response. We do not re-analyse other datasets; whether they regularize the lid, and how strongly, determines how far our numbers transfer.

The lid-edge singularity. Where a moving wall meets a stationary one at a right angle, the Stokes pressure behaves like νU/r\nu U/r at distance rr from the edge (the scraper problem; [47, 31, 43]). Inertia modifies this structure beyond r∼ν/Ur \sim\nu/U [16]. At 1283128^3 that length is about 13 cells at Re = 10 and less than one cell at Re ≥\ge 300, so the grid sees an inertia-modified corner at higher Reynolds numbers. Spectral benchmark solutions subtract the leading singular component analytically [6], three-dimensional reference solutions treat the edges with care [1, 21], and some direct and large-eddy simulations regularize the lid velocity because the singularity degrades convergence near the lid [23, 7]. Singular functions have also been built into physics-informed networks for corner singularities of the Poisson problem (Hu et al., 2024). Our reference correction plays a comparable role, but it is data-fitted and consistent with the solver’s own discretization. It removes the grid-level representation of the singularity, not the continuum singularity.

Reference-informed learning and resolution transfer. Reference Neural Operators learn a correction to a transported reference solution and note that the nearly constant density over most of the Airfoil-Euler domain can make whole-domain errors look small [10]. Multifidelity DeepONet learns a high-fidelity residual on top of a low-fidelity operator (Lu et al., 2022). Solver-in-the-loop training corrects a coarse numerical solver with a learned term (Um et al., 2020), CFD-GCN refines an upsampled coarse-mesh CFD solution, including pressure, with a graph network [12], and Hanna et al. [17] learn the local error of coarse-grid OpenFOAM solutions of the three-dimensional lid-driven cubic cavity. Coarse-to-fine correction in this very configuration is therefore not new; in all of these the correction is learned. Hybrid interpolation embeds CFD snapshots in a neural model (Sha et al., 2026). We instead freeze an immersed-geometry surrogate and add a deterministic residual from a body-free library; Table 1 compares the information each approach uses. Zero-shot super-resolution of learned operators is unreliable [4, 3] (Sakariadia et al. 2026), and directly executing our coarse surrogate at 1283128^3 indeed degrades accuracy (Section H). We therefore keep the learned field at its native 64364^3 scale.

MethodSolution information availableLearned vs. deterministic operations
Reference Neural Operators ([10])reference-domain solution, reference and query geometry, a constructed deformationtransport the reference; learn the query-minus-transported correction
Multifidelity DeepONet ((Lu et al., 2022))low- and high-fidelity training data; low-fidelity predictions as inputslearn a low-fidelity operator and a high-fidelity residual
Hybrid wind interpolation ([42])CFD snapshots for a fixed object and sampled inlet anglesinterpolation ansatz inside a trained neural model
Oursimmersed-geometry training fields; an offline body-free library; query geometry and Re⁡\operatorname{Re} only at inferencefreeze the surrogate; add the deterministic residual (I−LR2)B128(I-LR_2)B_{128}; center

Table 1. Information used by reference-informed approaches. These are methodological distinctions, not numerical comparisons.

Problem, data and evaluation

Flow configuration

Reference fields are steady, incompressible, laminar solutions of OpenFOAM v2412 simpleFoam (Weller et al. [53]) in the unit cube. The plane y=1y=1 moves with velocity (1,0,0)(1,0,0), and the other walls and the immersed bodies are no-slip. The kinematic viscosity is ν=1/Re⁡\nu=1/\operatorname{Re} with Re⁡∈{10,30,100,300,900}\operatorname{Re}\in\{10,30,100,300,900\}. On a uniform n3n^3 grid (n=64n=64 or 128128), cells whose centers have signed distance ≤0\le0 to the body are removed, so the immersed wall is voxelized. All test cases and all 17 reference-library cases use byte-identical discretization and linear-solver settings (Section D). Convergence requires residuals of 10−610^{-6} (pressure) and 10−710^{-7} (velocity) within 16,000 SIMPLE iterations. Pressure is kinematic and, since the lid speed and box size are one, it is measured in units of U2U^2. All fields are centered on the fluid mean:

ΠFq=m(q−1∣F∣∑i∈Fqi),m=1F.(1)\Pi_F q = m\left(q-\frac{1}{|F|}\sum_{i\in F}q_i\right), \qquad m=1_F. \tag*{(1)}

Cohorts

The coarse surrogate is trained on 1,620 64364^3 cases from 324 geometries of six families (sphere, cylinder, box, ellipsoid, torus, multi-body unions of two spheres) at the five Reynolds numbers. We evaluate at 1283128^3 on two disjoint test cohorts of three shape or composition categories unseen in training: capsules, heterogeneous pairs (sphere + box) and heterogeneous triplets (sphere + cylinder + box). Each family has four geometries, one per placement stratum (interior; near z=1z=1; near a random xx or yy face; near z=0z=0), evaluated at all five Reynolds numbers. Each cohort therefore has 60 cases and 12 independent geometry groups.

  • Cohort A (primary). Both the lifted surrogate and the reference-corrected predictor, the evaluator and the numerical criteria were sealed with hashes before the cohort’s CFD truth was generated.

  • Cohort B (second). This cohort was generated later, with a different generator seed, to test a different model. The reference-corrected predictor was that test’s sealed control arm, so its absolute scores on cohort B are prospective. The lifted arm was not sealed there. We reconstruct it deterministically from the sealed array by removing the correction (Equation (7)), which makes the cohort-B contrast POST HOC. Re-evaluating the sealed reference-corrected arrays reproduces its frozen cohort-B scores to 6×10−176\times10^{-17}.

Coarse-grid cohorts used during development are described in Section H and are not pooled with A or B.

Metrics and regions

For a region R\mathcal{R}, prediction and truth are centered independently on the same fluid mask and compared by

ER(p^,p)=∥ΠFp^−ΠFp∥2,R/∥ΠFp∥2,R.(2)E_{\mathcal{R}}(\hat{p},p)=\lVert\Pi_{\mathcal{F}}\hat{p}-\Pi_{\mathcal{F}}p\rVert_{2,\mathcal{R}}/\lVert\Pi_{\mathcal{F}}p\rVert_{2,\mathcal{R}}. \tag*{(2)}

Table 2 lists the six case metrics predeclared for cohort A; lower is better except for the recall RobjR_{\mathrm{obj}}.

MetricDefinition
EvolE_{\mathrm{vol}}Equation (2) on the full fluid volume F\mathcal{F}
EnearE_{\mathrm{near}}Equation (2) on the near-body band N\mathcal{N}
E∇E_{\nabla}Relative error of axiswise central differences where the center and both neighbours are fluid
EpkE_{\mathrm{pk}}Equation (2) on the truth’s top-5% ∣p∣|p| cells of F\mathcal{F} (global peaks)
EobjE_{\mathrm{obj}}Equation (2) on the truth’s top-5% ∣p∣|p| cells of N\mathcal{N}, excluding three cells at the outer walls (object peaks)
RobjR_{\mathrm{obj}}Fraction of those truth object-peak cells also in the prediction’s same-size top-5% set

Table 2. Predeclared case metrics (details and tie handling in Section A). All are dimensionless. The near-body band N={0<SDF≤0.2}\mathcal{N}=\{0<SDF\le0.2\} is 25.6 cells thick at 1283128^3 and overlaps the wall shell for most bodies; only EobjE_{\mathrm{obj}} and RobjR_{\mathrm{obj}} exclude cells at the outer walls.

Regions. For the spatial analysis we partition the fluid by cell distance to the box faces:

  • the wall shell S\mathcal{S}: cells within eight cells of any face;

  • the interior I=F∖S\mathcal{I}=\mathcal{F}\setminus\mathcal{S};

  • the lid-edge junction J⊂S\mathcal{J}\subset\mathcal{S}: cells within eight cells of both the lid and a stationary side wall. J\mathcal{J} has 30,720 cells (1.5% of the fluid): two blocks along the edges perpendicular to the lid motion, two along the parallel edges, and four vertex blocks.

The 8-cell width comes from a development diagnostic that predates cohort A (Section H). The split of the shell into junction and remaining wall classes was defined after cohort A had been evaluated, so all regional results are post hoc; Table 10 shows that the conclusions hold for junction widths from 1 to 16 cells. We also bin cells by Chebyshev cell distance to the nearest lid edge, i.e. max⁡(dlid,dside)\max(d_{\mathrm{lid}},d_{\mathrm{side}}) minimized over the side walls, and, within I\mathcal{I}, by distance to the body.

Width kk (cells)Cells (%)Truth energy share (%)Lifted sq. error share (%)Change inside JkJ_k (%)Change outside JkJ_k (%)
10.0218.588.4−85.0+6.8
20.1031.591.3−75.3+3.9
40.3849.393.2−69.7−9.0
81.4768.695.1−67.8−3.3
123.2077.995.7−67.3−0.8
165.4982.796.0−67.0−0.3

Table 10. Sensitivity to the junction width (cohort A, post hoc). JkJ_k contains the cells within kk cells of a lid edge (the bins of Figure 4a); changes are those of the cohort-mean regional error when adding the reference. “Outside” includes the rest of the wall shell; for k≤2k \le2 it contains the cell layers in which the reference over-corrects slightly.

Reference-relative body response. When an empty-cavity reference B128(Re)B_{128}(\mathrm{Re}) is available, define

d=ΠF[p−B128],d^=ΠF[p^−B128],Ebody,R=∥d^−d∥2,R∥d∥2,R.(3)d=\Pi_{\mathcal{F}[p-B_{128}],\qquad\hat{d}=\Pi_{\mathcal{F}[\hat{p}-B_{128}],\qquad E_{\mathrm{body},R}=\frac{\lVert\hat{d}-d\rVert_{2,R}}{\lVert d\rVert_{2,R}}.}} \tag*{(3)}

Both disturbances use the same reference, query fluid mask, region and full-fluid gauge. The reference is body-free before masking and centering on the query fluid. Since d^−d=ΠFp^−ΠFp\hat{d}-d=\Pi_{\mathcal{F}}\hat{p}-\Pi_{\mathcal{F}}p, only the denominator differs from Equation (2); this diagnostic is not independent evidence of accuracy or a decomposition into independent Navier–Stokes solutions. We average defined per-case ratios, not pooled squared norms. Zero disturbance energy makes the ratio undefined. Small positive energy can amplify relative errors, so we also report ∥d∥2,R2/∥ΠFp∥2,R2\lVert d\rVert_{2,R}^{2}/\lVert\Pi_{\mathcal{F}}p\rVert_{2,R}^{2} rather than clipping the score. The historical implementation used a 10−30010^{-300} safeguard on squared denominator energy; it was inactive in all 485 recorded cohort-A regions with nonzero truth energy (minimum disturbance energy 0.0152U40.0152U^{4}).

Energy decomposition. “Pressure energy” here abbreviates the squared L2L_{2} norm of centered pressure, not physical fluid energy. Let sR=∥ΠFp∥2,R2/∥ΠFp∥2,F2s_R=\lVert\Pi_{\mathcal{F}}p\rVert_{2,R}^{2}/\lVert\Pi_{\mathcal{F}}p\rVert_{2,\mathcal{F}}^{2} be a region’s share of that norm. For any exhaustive partition, with the common full-fluid centering,

Evol2=∑RsRER2.(4)E_{\mathrm{vol}}^{2}=\sum_R s_R E_R^{2}. \tag*{(4)}

The weights sRs_R are independent of the model, but its error-energy shares wR=sRER2/Evol2w_R=s_RE_R^{2}/E_{\mathrm{vol}}^{2} are not (for Evol>0E_{\mathrm{vol}}>0). Thus large junction truth energy does not imply that every prediction’s error is concentrated there. With sJ≈0.69s_{\mathcal{J}}\approx0.69 and sI≈0.13s_{\mathcal{I}}\approx0.13 (Table 7), an interior error of 10% contributes at most 0.13×10%≈3.7%\sqrt{0.13}\times10\%\approx3.7\% to EvolE_{\mathrm{vol}}.

Region RRCellsTruth energyLiftedNo-learningOursCohort A +ReferenceCohort A +LearnedCohort B +Reference
Junction J\mathcal{J}1.568.624.47.97.8−67.8−0.3−67.7
Rest of wall shell S∖JS \setminus\mathcal{J}31.718.05.79.15.3−6.7−42.2—
Interior I\mathcal{I}66.913.46.917.36.9+0.2−59.9+0.2
Near body12.32.110.040.010.00.0−75.0—
Elsewhere54.611.46.310.76.3+0.3−40.6—

Table 7. Regional decomposition (cohort A unless marked; POST HOC). Cell and truth-energy shares are case means. ERE_R applies Equation (2) on region RR with the canonical full-fluid centering. The last three columns give the change of the cohort-mean regional error: when adding the reference (lifted surrogate → ours), when adding the learned term (no-learning control → ours; RETROSPECTIVE), and when adding the reference on cohort B. “Near body” is N∩I\mathcal{N} \cap\mathcal{I}; “–”, not computed for cohort B.

Statistics. The geometry is the independent unit (n=12n=12 per cohort), and the five Reynolds numbers of a geometry are repeated measures. We report:

  • cohort-mean error changes, i.e. the ratio of arithmetic means; recall uses the mean paired geometry-level difference;

  • 95% tt intervals for the geometric-mean error ratio over geometries, a different estimand, so an interval need not contain the change of means; recall instead uses paired geometry-level differences and their tt interval;

  • per-geometry ranges and leave-one-geometry-out ranges;

  • two-sided sign-flip tests enumerating all 2122^{12} patterns of geometry means. Enumeration is exact; inferential validity additionally requires sign symmetry under the null, not merely independent geometries.

For cohort A, the following decision rule was predeclared (Section A.1): at least 10% mean reduction of EvolE_{\mathrm{vol}}, E∇E_{\nabla} and EpkE_{\mathrm{pk}}; at most 2% mean increase of EnearE_{\mathrm{near}} and EobjE_{\mathrm{obj}}; at most 0.01 loss in RobjR_{\mathrm{obj}}. Reading those tolerances as non-inferiority margins is our post hoc choice. The resulting log-ratio intervals concern geometric-mean ratios, not the predeclared arithmetic-mean decision rule; they do not certify non-inferiority for that rule. Everything else is descriptive.

Method: a frozen surrogate plus a body-free native-grid reference

Figure 2 summarizes the predictor.

Figure 2. Reference-corrected lift. Box outlines: blue, learned; red, CFD-derived and body-free; grey, deterministic operators, including the final assembly (bottom). The lifted surrogate omits the right branch. The no-learning control sets cθ=0c_{\theta}=0 inside p^64\hat{p}_{64}, keeping the coarse background b64b_{64}. No query pressure and no query-time CFD enter the prediction.

Frozen coarse surrogate. The learned component is a 64364^{3} surface-conditioned network with 3.15M parameters, developed and frozen before this study. Section B specifies it fully; its individual modules are not the subject of this paper.

  • Output. It predicts a centered correction cθc_{\theta} to a Reynolds-only background b64(Re)b_{64}(\mathrm{Re}), the family-balanced mean of training pressures, and outputs p^64=ΠF64[b64(Re)+cθ]\hat{p}_{64}=\Pi_{\mathcal{F}_{64}}[b_{64}(\mathrm{Re})+c_{\theta}].

  • Architecture. The volume pathway is a masked U-Net hierarchy 64→32→16→8→464\to32\to16\to8\to4 (Ronneberger et al., [39]; Çiçek et al., [11]). Nine encoder and decoder stages exchange features with a component-safe graph on the immersed surface and with an area-weighted global surface token.

  • Inputs. Six dense channels: the SDF, three coordinates, the outer-wall distance and a scaled log Re, all zeroed in solid. No velocity, geometry identifier or pressure enters.

  • Training. 40,500 single-case AdamW updates on the 1,620 training cases, with a loss on the complete centered field, the near band, object peaks and gradients. The terminal checkpoint was fixed in advance.

Reference correction. Let B128(Re)B_{128}(\mathrm{Re}) be a centered empty-cavity pressure. It is reconstructed from a rank-8 SVD of 17 native-grid empty-cavity solutions at Re∈[10,997]\mathrm{Re} \in[10,997], with modal coefficients interpolated piecewise-cubically in log⁡10Re\log_{10}\mathrm{Re} (Section C). Let R2R_2 be 2×2×22 \times2 \times2 block averaging and LL the corresponding lift. The correction

δB128(Re)=B128(Re)−LR2B128(Re)(5)\delta B_{128}(\mathrm{Re}) = B_{128}(\mathrm{Re}) - L R_2 B_{128}(\mathrm{Re}) \tag*{(5)}

is the restriction–prolongation residual of the reference: the part of the empty-cavity field that block averaging followed by the lift discards. It is not an optimal representation residual: an attained numerical least-squares fit of B128B_{128} by LL applied to a 64364^3 field, plus a constant, leaves 17.1% of its norm at Re=10\mathrm{Re}=10 and 8.4% at Re=900\mathrm{Re}=900, against 21.2% and 10.6% for δB128\delta B_{128} (Section E). The predictor is

p^128(g,Re)=ΠF[Lmp^64(g,Re)+δB128(Re)].(6)\hat{p}_{128}(g,\mathrm{Re}) = \Pi_F\left[L_m\hat{p}_{64}(g,\mathrm{Re})+\delta B_{128}(\mathrm{Re})\right]. \tag*{(6)}

Here LmL_m is a fluid-mask-normalized trilinear transfer that avoids mixing solid fill into fluid values (Section A). The query geometry gg supplies both the 64364^3 and the 1283128^3 fluid masks. The predictor accepts only the five training Reynolds numbers, because b64b_{64} is a lookup table. At its 17 fit points the rank-8 reference reproduces the library fields to a relative error of at most 3.5×10−43.5 \times10^{-4} (5.5×10−65.5 \times10^{-6} at Re=10\mathrm{Re}=10, 4.8×10−54.8 \times10^{-5} at Re=100\mathrm{Re}=100). For Re=30,300,900\mathrm{Re}=30,300,900 it is interpolated; on 16 held-out Reynolds numbers between 11 and 864 the reference reproduces native empty-cavity CFD with mean relative error 0.072% and worst 0.198% (Table 3).

Held-out Re (16 values, 11–864)Meanp90WorstRe 11–37 / 48–205 / 274–864
EvolE_{\mathrm{vol}} of B128(Re)B_{128}(\mathrm{Re}) (%)0.0720.1640.1980.039 / 0.051 / 0.131
E∇E_{\nabla} of B128(Re)B_{128}(\mathrm{Re}) (%)0.0710.1620.244–
1−ρ1-\rho, field correlation (10−610^{-6})0.371.341.87

Table 3. Held-out-Reynolds fidelity of the empty-cavity reference (obstacle-free; 16 held-out values). Relative errors in %; 1−ρ1-\rho in units of 10−610^{-6} (ρ\rho: correlation between the reference and the held-out CFD field). The last column splits the 16 values into low, middle and high Reynolds numbers (5/6/5 values).

Arms. Four compositions differ only in whether the learned correction cθc_\theta and the reference correction δB128\delta B_{128} are present:

  • the lifted surrogate, ΠFLmp^64\Pi_F L_m\hat{p}_{64}, which omits the reference;

  • ours, Equation (6);

  • the no-learning control, ΠF[LmΠF64b64+δB128]\Pi_F[L_m\Pi_{F_{64}}b_{64}+\delta B_{128}], which keeps every deterministic term;

  • background only, ΠFLmΠF64b64\Pi_F L_m\Pi_{F_{64}}b_{64}.

Because centering on a fixed mask is linear,

p^128−p^128lift=ΠFδB128(Re),(7)\hat{p}_{128}-\hat{p}^{\mathrm{lift}}_{128}=\Pi_F\delta B_{128}(\mathrm{Re}), \tag*{(7)}

so the prospective contrast isolates exactly the reference correction.

Results

Prospective effect of the reference

Table 4 and Figure 3a give the result.

Evol↓E_{\mathrm{vol}} \downarrowEnear↓E_{\mathrm{near}} \downarrowE∇↓E_{\nabla} \downarrowEpk↓E_{\mathrm{pk}} \downarrowEobj↓E_{\mathrm{obj}} \downarrowRobj↑R_{\mathrm{obj}} \uparrow
Cohort A (primary): 60 cases, 12 unseen geometries
Lifted surrogate20.610.153.521.88.486.0
Reference-corrected (ours)7.59.612.57.28.486.0
Change−63.5%−5.5%−76.6%−66.9%+0.1%+0.01 pp
95% CI (geometry)[−64.4, −62.6][−15.8, +6.6][−76.8, −76.3][−67.2, −66.6][−0.6, +1.1][−0.02, +0.03]
Geometries improved12/127/1212/1212/128/128/12
Cohort B (second): 60 cases, 12 unseen geometries
Lifted surrogate20.69.853.521.99.883.3
Reference-corrected (ours)7.49.812.57.39.783.3
Change−64.2%−0.1%−76.7%−66.8%−0.1%+0.01 pp
95% CI (geometry)[−64.7, −63.6][−0.3, 0.0][−76.8, −76.5][−67.0, −66.6][−0.2, 0.0][−0.01, +0.03]
Geometries improved12/1210/1212/1212/1210/127/12

Table 4. Adding the body-free reference to the frozen surrogate (cohort-mean relative errors, %). Cohort A is PROSPECTIVE: both arms were sealed before truth. In cohort B our predictor was sealed, and the lifted arm is its deterministic reconstruction (POST HOC; storage-level agreement checked on five cohort-A cases). Error changes are ratios of cohort means; recall changes are percentage-point differences. For errors, the 95% tt intervals concern the geometric-mean ratio over the 12 geometries; for recall they concern the mean paired geometry-level difference. “Improved” counts geometries with a lower (recall: higher) mean. Bold marks cohort-A metrics whose reduction exceeds the predeclared 10% requirement; the requirement applied to cohort A only.

Figure 3. Metric-level effect of each component: signed change of the cohort-mean relative error, negative is better, as in Table 4. a, Adding the reference to the frozen surrogate (lifted surrogate → ours; cohort A PROSPECTIVE, cohort B POST HOC). Bars are ratios of arithmetic cohort means; no confidence intervals are attached to them. Table 4 separately reports intervals for geometry-level geometric-mean ratios, a different estimand. Each cohort has 12 independent geometry groups; capsule-02 dominates the cohort-A near-body mean change, and near-body non-inferiority is not established. b, Adding the learned correction (no-learning control → ours; cohort A, RETROSPECTIVE; cohort means only). Peak-region recall changes by +0.01 points in (a) and by +24.1 points in (b).

  • Global metrics. Adding the reference lowers the global error from 20.6% to 7.5% (−63.5%), the gradient error from 53.5% to 12.5% (−76.6%) and the global peak error from 21.8% to 7.2% (−66.9%).

All three decrease in every one of the 12 geometries (enumerated sign-flip p=0.0005p = 0.0005, the minimum attainable under the sign-symmetry null). The geometry-level reductions span 60.6–65.5%, 75.4–76.9% and 66.1–67.6%. The effect shrinks with Reynolds number: EvolE_{\mathrm{vol}} falls by 67% at Re=10\mathrm{Re} = 10 but by 50% at Re=900\mathrm{Re} = 900 (Table 6).

CohortReEvolE_{\mathrm{vol}} (%) LiftedEvolE_{\mathrm{vol}} (%) OursChangeEnearE_{\mathrm{near}}EobjE_{\mathrm{obj}}RobjR_{\mathrm{obj}}Worst EobjE_{\mathrm{obj}}Worst RobjR_{\mathrm{obj}}
A1024.88.2−66.8-66.88.88.692.217.186.8
A3024.68.2−66.9-66.98.78.291.914.185.7
A10022.87.6−66.7-66.79.68.988.015.975.0
A30017.46.9−60.5-60.510.77.487.813.581.0
A90013.56.8−49.5-49.510.09.069.922.841.5
B1025.08.2−67.0-67.010.611.890.319.784.8
B3024.88.2−67.0-67.010.211.289.019.580.7
B10022.97.5−67.2-67.210.711.186.316.669.8
B30017.16.3−63.1-63.19.87.684.411.465.1
B90013.46.7−49.7-49.77.67.066.319.043.6

Table 6. Reynolds-number groups (12 geometries each; the same geometries recur across rows). Errors and recall are cohort means in %; “change” is the relative change of EvolE_{\mathrm{vol}} when adding the reference; “worst” is the largest error or smallest recall among the 12 cases.

  • Object-local metrics. They do not move: EobjE_{\mathrm{obj}} changes by +0.08% and RobjR_{\mathrm{obj}} is unchanged (less than one cell of the top-5% set on average). Their one-sided 95% bounds (+0.95% and −0.014 points) lie inside the predeclared tolerances. For EnearE_{\mathrm{near}} the bound is +4.28%, above the 2% tolerance, because the capsule-02 outlier widens the interval (Table 21); we therefore do not claim non-inferiority for EnearE_{\mathrm{near}} in cohort A.

CohortMetricChange of means (%)Median geometry (%)LOGO range (%)One-sided 95% bound (%)Geometries improvedSign-flip pp
AEvolE_{\mathrm{vol}}−63.46−63.16[−63.72, −63.28]−62.7412/120.0005
AEnearE_{\mathrm{near}}−5.53−0.08[−6.11, −0.05]+4.287/120.2168
AE∇E_{\nabla}−76.59−76.70[−76.69, −76.55]−76.3912/120.0005
AEpkE_{\mathrm{pk}}−66.89−66.93[−66.96, −66.82]−66.6612/120.0005
AEobjE_{\mathrm{obj}}+0.08−0.09[−0.12, +0.13]+0.958/120.9854
ARobjR_{\mathrm{obj}} (pp)+0.006––−0.014{sup}†`8/120.6108
BEvolE_{\mathrm{vol}}−64.16−64.49[−64.33, −64.07]−63.7112/120.0005
BEnearE_{\mathrm{near}}−0.13−0.17[−0.17, −0.10]−0.0410/120.0513
BE∇E_{\nabla}−76.69−76.73[−76.74, −76.67]−76.5712/120.0005
BEpkE_{\mathrm{pk}}−66.81−66.84[−66.86, −66.76]−66.6312/120.0005
BEobjE_{\mathrm{obj}}−0.10−0.07[−0.12, −0.08]−0.0410/120.0049
BRobjR_{\mathrm{obj}} (pp)+0.010––−0.007†7/120.3232

Table 21. Geometry-level statistics for adding the reference (n=12n = 12 geometries per cohort; cohort-B contrast post hoc). LOGO: range of the arithmetic-mean change over leave-one-geometry-out subsets. The one-sided bound is the upper 95% bound on the geometric-mean ratio of geometry-mean errors, expressed as a relative increase (tt on log ratios); †lower 95% bound on the mean paired recall change. Non-inferiority margins are +2% for EnearE_{\mathrm{near}}, EobjE_{\mathrm{obj}} and −1 pp for RobjR_{\mathrm{obj}}. In cohort A, the EnearE_{\mathrm{near}} bound exceeds its margin because of the favourable capsule-02 outlier (§ 5.5); excluding it post hoc, the bound is below +0.1%, which does not establish non-inferiority for the full cohort.

  • Decision rule. It is met in 11 of 12 geometries, all three families and all five Reynolds numbers. The exception is capsule-02: its near-body error drops by 47% while its object-peak error rises by 4.6%. This single geometry produces the −5.5% mean change of EnearE_{\mathrm{near}}; the median geometry change is −0.08% (§ 5.5).

The diagnostic comparison in Table 5 sharpens the primary result: the body-free reference alone scores 8.9% globally but 43.9% near the bodies, whereas the reference-corrected predictor scores 7.5% and 9.6%. Rebasing lowers the global score further to 3.2% without comparable local gain (Table 7). Reference alone and rebase are RETROSPECTIVE, not additional prospective tests; neither is the matched no-learning arm.

ArmEvol↓E_{\mathrm{vol}}\downarrowEnear↓E_{\mathrm{near}}\downarrowE∇↓E_{\nabla}\downarrowEpk↓E_{\mathrm{pk}}\downarrowEobj↓E_{\mathrm{obj}}\downarrowRobj↑R_{\mathrm{obj}}\uparrowEJE_JEIE_IEfaceE_{\mathrm{face}}
Cohort A
Lifted surrogate20.610.153.521.88.486.024.46.973.1
Reference-corrected (ours)7.59.612.57.28.486.07.86.910.1
Reference alone, ΠFB128(Re)\Pi_F B_{128}(\mathrm{Re})8.943.93.63.339.355.81.420.53.3
Rebased on the native reference3.29.42.51.28.286.00.76.72.6
Cohort B
Lifted surrogate20.69.853.521.99.883.324.46.273.1
Reference-corrected (ours)7.49.812.57.39.783.37.96.210.0
Reference alone, ΠFB128\Pi_F B_{128}7.239.83.22.438.752.11.216.82.9
Rebased on the native reference2.89.72.31.29.683.30.75.92.4

Table 5. Diagnostic baselines (cohort-mean relative errors and recall, %). Reference alone and rebase are retrospective; the original cohort-A comparison remains prospective, and the cohort-B lifted contrast is post hoc (Table 4). The rebased surrogate is Equation (8); the reference alone is the body-free ΠFB128(Re)\Pi_F B_{128}(\mathrm{Re}). EJE_J and EIE_I are the junction and interior errors of Table 7; EfaceE_{\mathrm{face}} is the relative error of differences across interior faces between fluid cells.

In cohort B, the global, gradient and peak reductions are −64.2%, −76.7% and −66.8%, with all 12 geometries improved. EnearE_{\mathrm{near}} and EobjE_{\mathrm{obj}} change by less than 0.5% and RobjR_{\mathrm{obj}} by less than 0.1 points in every geometry. The prospective absolute scores of our predictor on cohort B are within 0.3 percentage points of cohort A for the global, gradient, peak and near-body metrics. Object-local scores are somewhat worse (EobjE_{\mathrm{obj}} 9.7% vs. 8.4%, RobjR_{\mathrm{obj}} 83.3% vs. 86.0%).

Where the error lives

Table 7 decomposes each arm’s error on the region partition of § 3.3. The partition reproduces every case’s EvolE_{\mathrm{vol}} through Equation (4) to 8×10−168 \times10^{-16}.

  • Junction. It is 1.5% of the fluid but holds 68.6% of the centered truth energy, and 95% of the lifted surrogate’s squared error. The share is nearly independent of Reynolds number (66.6–69.8%).

Adding the reference lowers the junction error from 24.4% to 7.8% (−67.8-67.8%; interval [−68.1-68.1, −67.5-67.5]%; 12 of 12 geometries).

  • Interior. The interior error changes by +0.16+0.16% (interval [−0.14-0.14, +0.35+0.35]%). Only 3 of 12 geometries improve, but the interval includes zero; the data do not establish a consistent direction. The reference has a small interior tail.

  • Cohort B. The junction change is −67.7-67.7% and the interior change +0.16+0.16%.

  • Gradient and peak metrics. They are junction effects too (Table 8). The wall shell carries 98.7% of the truth’s gradient energy, and interior E∇E_{\nabla} changes by +0.09+0.09%. The 17% of global peak cells that lie in J\mathcal{J} carry 98.9% of the lifted surrogate’s peak error.

MetricRegionTruth share (%)Share of sq. error (%) LiftedShare of sq. error (%) OursRegional error (%) LiftedRegional error (%) OursMedian ratio ours/lifted
E∇E_{\nabla}Wall shell S\mathcal{S}98.799.7—53.812.3—
Junction J\mathcal{J}94.599.490.1———
Interior I\mathcal{I}1.30.35.322.622.6—
EpkE_{\mathrm{pk}}Junction J\mathcal{J}17.2*98.992.5——0.32
Interior I\mathcal{I}32.1*0.32.2——1.01

Table 8. Gradient and global-peak errors by region (cohort A, POST HOC). *Share of the global top-5% peak support rather than of energy. The last column is the median over cases of our regional error divided by the lifted one; for EpkE_{\mathrm{pk}} the J\mathcal{J} mask is the audit’s lid-edge band, whose per-case errors coincide with those on J\mathcal{J}.

  • A blind spot of E∇E_{\nabla}. Central differences over two cells do not see the grid-scale checkerboard mode. For the lifted surrogate, the residual after block averaging and trilinear re-extension has 55% of the original error energy (2.6% for ours). This operator-residual ratio is not a frequency-band or checkerboard energy fraction: the transfer is not an orthogonal spectral projection. The relative error of face differences between neighbouring fluid cells, which sees this mode, is 73.1% for the lifted surrogate and 10.1% for ours (Table 5). The ranking does not change, but E∇E_{\nabla} understates the error of lifted fields.

Anatomy of the junction. Table 9 splits J\mathcal{J} as the physics suggests. The two edge blocks perpendicular to the lid motion, where the lid scrapes against a side wall, hold 59.9% of the truth energy in 0.69% of the fluid; the four vertex blocks add 7.9%. The two edges parallel to the motion hold only 0.8%, consistent with a bounded pressure there. Figure 4a resolves the junction by distance to the nearest lid edge. The single layer of cells on the edges holds 18.5% of the truth energy and, on average over cases, 88% of the lifted surrogate’s squared error. There the lifted local RMS error is about 12.5×12.5\times the case’s truth RMS, and ours is about 1.9×1.9\times. In the next two cell layers the reference over-corrects slightly; from 16 cells onward the lifted and reference-corrected curves agree to within about 1%. In the interior (Figure 4b) the reference changes nothing at any distance from the body, whereas removing the learned term raises the local error by a factor of 1.4 to 4.9.

Figure 4. Error profiles (cohort A, 60 cases; post hoc; the two arms without the learned term are retrospective). Local RMS error per distance bin as a percentage of the case’s full-fluid truth RMS, averaged over cases (log scale). Bins widen with distance but are drawn at equal spacing. The upper curves are per-bin RMS values, not quantities that accumulate linearly with the number of cells; their values can still change with bin boundaries and widths; the lower strips give each bin’s share of truth energy, which grows with the number of cells in the bin. “Background only” is the lifted Reynolds-only background (§ 4), with neither the learned correction nor the reference; away from the lid edges it coincides with the no-learning control. a, Chebyshev cell distance to the nearest lid edge; the dotted line bounds the 8-cell junction. Near the edges our predictor (black) follows the no-learning control (red); far from them it follows the lifted surrogate (blue). Markers follow Figure 1c: open circle, lifted; triangle, no-learning control; filled dot, ours. b, Cell distance to the immersed surface within the interior I\mathcal{I}. The lifted and reference-corrected curves agree to within 0.2% (the blue curve lies under the black one), and the arms without the learned term are 1.41.4–4.9×4.9\times worse.

SubregionCellsTruth energy share (%)LiftedNo-learningOursΠFB128(Re)\Pi_F B_{128}(\mathrm{Re}) alone
Perpendicular edges (x=0,1x=0,1)14,33659.924.47.77.71.1
Parallel edges (z=0,1z=0,1)14,3360.85.77.55.48.1
Vertex blocks2,0487.925.48.98.91.2
Junction J\mathcal{J} (all)30,72068.624.47.97.81.4

Table 9. Junction anatomy (cohort A, post hoc; case means; cells are counts, all other values in %). Blocks along the two edges perpendicular to the lid motion (excluding the vertex blocks), along the two parallel edges, and the four 838^3 vertex blocks. The last column evaluates the body-free reference ΠFB128(Re)\Pi_F B_{128}(\mathrm{Re}) alone against the truth.

What the reference corrects. Three observations identify its target.

  • It does not depend on the body. The body-free field ΠFB128\Pi_F B_{128} alone matches the truth in J\mathcal{J} to 1.4%, and in the perpendicular-edge blocks to 1.1% (Table 9). This holds at every Reynolds number (1.2–1.5%) and stays below 3.7% in every case; the largest geometry mean, 2.8%, belongs to capsule-02, whose near-body band reaches into the junction.

  • Its settings match the truth. The reference library and the test truth use byte-identical discretization and solver settings (Section D).

  • The estimated representation residual is concentrated at the lid edges. A numerical full-fluid least-squares fit of LmL_m applied to a 64364^3 field leaves 13.8% of the truth norm on average. That global fit’s regional residuals are 16.6% in J\mathcal{J}, 2.9% in the interior near-body band and 1.0% in the interior; they are not independently optimized regional lower bounds. Of the fitted residual energy, 99.9% lies in J\mathcal{J} (Section E).

The reduction in EvolE_{\mathrm{vol}} is therefore agreement with the solver’s own lid-edge values, not an improvement of the pressure response to the body. The 69% junction share is also a property of this grid rather than of the continuum. Near a 1/r1/r singularity the energy grows logarithmically with refinement, and in the empty cavity the share of a band of fixed physical width grows from 64364^3 to 2563256^3 while that of a band of fixed cell count falls (§ 5.7). The correction removes about two thirds of the junction error. Yet the junction still holds 75% of our predictor’s remaining squared error, because the lifted coarse field, not the reference, sets the error there: 7.8% for the reference-corrected predictor versus 1.4% for the reference alone. § 5.3 tests the obvious consequence.

A frozen audit label. Before this regional analysis we froze a separate audit with a descriptive rule. It labels the reference gain “boundary dominated” only if every one of its inputs is below a small limit. That rule returns “spatially mixed”. Its only failing input is the near-body squared-error change pooled over geometries at each Reynolds number (−22% to −61%), and capsule-02’s junction cells dominate it. Every other input is below 10−410^{-4} against a limit of 0.05. Excluding capsule-02, as a post hoc sensitivity analysis, the pooled near-body changes lie within ±0.5%\pm0.5\% at every Reynolds number. We report the frozen label unchanged (§ 17).

Diagnostic baselines: the reference alone and a rebased surrogate

The correction δB128\delta B_{128} adds only what block averaging removes from B128B_{128}. The surrogate’s own coarse representation of the lid edges, learned from 64364^3 training data, stays in the prediction. A rebased surrogate replaces it by the native reference:

p^128rebase=ΠF[B128+Lm(p^64−B64)]=p^128lift+ΠF[B128−LmB64].(8)\hat{p}_{128}^{\mathrm{rebase}} = \Pi_F\left[B_{128} + L_m(\hat{p}_{64} - B_{64})\right] = \hat{p}_{128}^{\mathrm{lift}} + \Pi_F\left[B_{128} - L_m B_{64}\right]. \tag*{(8)}

Here B64(Re)B_{64}(\mathrm{Re}) is an empty-cavity CFD solution on the 64364^3 grid with the settings of the reference library. Like B128B_{128} it is body-free and computed offline, once per Reynolds number. The rebased surrogate therefore differs from ours only in its body-independent term. We also evaluate the reference alone, ΠFB128\Pi_{\mathcal{F}}B_{128}, which contains no learned body-conditioned response; it still uses the query fluid mask for centering and evaluation. Both are retrospective diagnostics, designed after both cohorts had been evaluated; cohort B uses the reconstructed lifted arm.

Table 5 gives three results, each replicated on cohort B.

  • Similar global scores can hide different body accuracy. The reference alone has a global error of 8.9% (A) and 7.2% (B), close to ours (7.5% and 7.4%), while its near-body error is 44% and 40%. It also beats the body-aware lifted model on all three global metrics in cohort A (volume: 8.9% versus 20.6%; gradient: 3.6% versus 53.5%; global peaks: 3.3% versus 21.8%), despite its much worse near-body error (43.9% versus 10.1%). These are retrospective diagnostic comparisons, not an additional prospective superiority test.

  • Rebasing lowers every global metric further. EvolE_{\mathrm{vol}} falls from 7.5% to 3.2% (A) and from 7.4% to 2.8% (B), and the junction error from 7.8% to 0.7%. The per-geometry ratio of EvolE_{\mathrm{vol}} to ours is 0.29–0.60 (A) and 0.31–0.46 (B).

  • Cohort-mean body metrics change little. EnearE_{\mathrm{near}}, EobjE_{\mathrm{obj}}, RobjR_{\mathrm{obj}} and the interior error move by at most 0.33 points, and the interior near-body error by at most 0.1 points. The largest per-geometry change of EnearE_{\mathrm{near}}, −30%, is again capsule-02.

Our composition is thus not the best way to use the reference, and we do not claim it is. Its role in this paper is that of a controlled intervention: Equation (7) isolates one body-independent term, whose effect can be located exactly. The rebased surrogate shows that most of the remaining global error of ours can also be removed by a body-independent term, with little change in the cohort-mean body metrics. Importantly, the junction’s share of error is model-dependent: its case-mean wJw_J in cohort A is 95.1% for lifted, 75.4% for ours, but only 1.9% for reference alone and 4.0% for rebase. These post hoc shares are computed per case before averaging, not by multiplying mean truth weights and mean errors. A large truth-energy share therefore does not force a predictor’s error to lie in the junction.

The learned term carries the body response

The no-learning control keeps the coarse background, both masks, the transfer, the reference correction and the centering, and removes only the learned correction (Table 11; retrospective). It was constructed without reading any pressure truth, but after cohort A had been inspected.

Arm (cohort A)StatusEvol↓E_{\mathrm{vol}}\downarrowEnear↓E_{\mathrm{near}}\downarrowE∇↓E_{\nabla}\downarrowEpk↓E_{\mathrm{pk}}\downarrowEobj↓E_{\mathrm{obj}}\downarrowRobj↑R_{\mathrm{obj}}\uparrow
Background onlyR21.838.353.622.136.461.9
Lifted surrogateP20.610.153.521.88.486.0
No-learning controlR10.138.012.97.936.461.8
Reference-corrected (ours)P7.59.612.57.28.486.0
Ours, 3 retrained seeds (mean)R7.49.612.57.28.585.6
Matched residual FNO, 3 seeds (mean)R7.912.612.77.311.481.7

Table 11. Four compositions and two retrospective comparisons (cohort A, cohort-mean relative errors in %). P: prospective; R: retrospective, an arm or retraining created after the cohort had been inspected. The seed replication and the matched FNO use the same data, inputs, background, 40,500 updates, lift, reference and evaluator. The FNO has 96.8% of the surrogate’s real-valued parameters.

  • Metric level. Relative to the control, our predictor lowers EnearE_{\mathrm{near}} by 74.8%, EobjE_{\mathrm{obj}} by 76.9% and EvolE_{\mathrm{vol}} by 25.0%, and it raises RobjR_{\mathrm{obj}} by 24.1 points. EnearE_{\mathrm{near}} and EobjE_{\mathrm{obj}} improve in all 12 geometries, in every family and at every Reynolds number.

  • Region level. The learned term mirrors the reference (Table 7, Figure 5). It removes 59.9% of the interior error and 75.0% of the interior near-body error, but only 0.3% of the junction error. The separation holds in every geometry: each component moves the other’s region by at most 5.3% (Figure 5).

Figure 5. Separable roles in every geometry (cohort A; post hoc regional analysis, retrospective no-learning control). Per-geometry mean regional error in the lid-edge junction (a) and in the interior (b): lifted surrogate (upper lane), ours (middle) and no-learning control (lower lane). Horizontal rules separate the capsule, pair and triplet families. In every geometry, adding the reference changes the junction error by −68.5% to −67.0% and the interior error by −0.8% to +0.6%; adding the learned term changes the interior error by −68.0% to −53.2% and the junction error by −5.3% to +1.4%.

  • Body-local is not body-induced. Near the bodies most of the pressure is still cavity flow. The reference-relative disturbance dd of Equation (3) carries 23.5% of the truth energy in the interior near-body band, 19.3% on the object-peak support, 4.9% in the interior and 0.03% in J\mathcal{J} (means of case-level energy ratios). Its smallest energy ratio in the interior near-body band is 0.00582. Measured against ∥d∥\lVert d\rVert instead of ∥ΠFp∥\lVert\Pi_{\mathcal{F}}p\rVert, the interior near-body error is 24.6% for ours, 24.7% for the lifted surrogate and 88.8% for the no-learning control; on the object-peak support it is 26.0%. In the interior near-body band, the ratio of the two cohort-mean error normalizations is about 2.5; this is not a per-case conversion factor. The correction barely changes this reference-relative body-response diagnostic.

Denote the four arms’ errors by e00e_{00} (background only), e10=e00+ℓe_{10}=e_{00}+\ell (learned term added), e01=e00+re_{01}=e_{00}+r (reference added) and e11=e00+ℓ+re_{11}=e_{00}+\ell+r. Here ℓ\ell is the lifted learned correction and r=ΠFδB128r=\Pi_{\mathcal{F}}\delta B_{128}. Because every operator is linear for fixed masks, the interaction of squared relative errors is

(E002−E102)−(E012−E112)=2⟨ℓ,r⟩∥ΠFp∥2,(9)(E_{00}^{2}-E_{10}^{2})-(E_{01}^{2}-E_{11}^{2})=\frac{2\langle\ell,r\rangle}{\lVert\Pi_F p\rVert^{2}}, \tag*{(9)}

whose unnormalized numerator, on fixed support, does not depend on the truth. The normalized interaction does depend on the truth energy and, on truth-selected peak regions, on that support. Its case mean is +9.7×10−6+9.7 \times10^{-6}, 0.2% of the smaller main effect (per-case range −1.9×10−4-1.9 \times10^{-4} to +6.7×10−5+6.7 \times10^{-5}). The small interaction shows near-additivity of squared losses in these cases, not disjoint spatial support: overlapping zero-mean vectors can have zero inner product. Regional improvements provide separate evidence about where the corrections help. The identity also holds on every region with the regional truth energy in the denominator. There the interaction is at most 2.3×10−32.3 \times10^{-3} in any case and region, including the object-peak support, and at most 1.2×10−31.2 \times10^{-3} for capsule-02. The square of EobjE_{\mathrm{obj}} also obeys this identity because its truth-selected support is fixed across the four arms. Recall RobjR_{\mathrm{obj}} does not: its prediction-selected support varies. Capsule-02’s object-peak error can worsen through alignment of the added reference with the existing error, without a change of truth-selected support (§ 5.5).

A matched spectral comparator. Is this local accuracy specific to the learned surrogate? We trained a residual masked FNO (Li et al., 2021) under the same contract (RETROSPECTIVE). It uses the same 1,620 training cases, deployment-visible inputs, Reynolds-only background and 40,500 updates, three predeclared seeds, and the same lift, reference and evaluator, so the comparison is between two learned surrogates inside the same composition. The FNO has four spectral blocks, width 21 and 636^3 retained modes out of 32 per axis, for 3.05M real-valued parameters (96.8% of the surrogate’s). Its optimizer, learning rate and schedule were inherited from the surrogate and not tuned. The mode count was kept from an earlier FNO baseline, and the width was set to match capacity. The FNO reaches similar global accuracy but is worse locally (Table 11). Averaged over seeds, the reference-corrected surrogate has 24.3% lower EnearE_{\mathrm{near}}, 25.6% lower EobjE_{\mathrm{obj}} and 3.9 points higher recall. For every seed it passes all nine checks of a predeclared symmetric superiority rule, and no FNO seed passes the reverse rule (§ 17.§ 7). The strong low-pass restriction of 6 modes is a plausible reason for the local deficit. The result is evidence against one untuned FNO configuration, not against the FNO family.

Seed sensitivity. We retrained the surrogate from scratch with three predeclared seeds under the same configuration (RETROSPECTIVE). The cohort means changed by at most 0.08 points (EvolE_{\mathrm{vol}}), 0.34 (EnearE_{\mathrm{near}}), 0.23 (EobjE_{\mathrm{obj}}) and 0.37 (RobjR_{\mathrm{obj}}). Including the originally sealed run, the four-run ranges are 0.18, 0.34, 0.23 and 0.51 points (Table 12). The tails vary more: the seed ranges of the 90th percentiles of EnearE_{\mathrm{near}} and EobjE_{\mathrm{obj}} are 1.1 and 0.9 points. An unblinded AI-agent visual review of all 24 fixed-scale panels (8 cases × 3 seeds) reported no new artifacts. The effects of Sections 5.1 and 5.4 are one to two orders of magnitude larger than these ranges; the 5.5% mean change of EnearE_{\mathrm{near}} in cohort A is not.

Run (cohort A)Evol↓E_{\mathrm{vol}} \downarrowEnear↓E_{\mathrm{near}} \downarrowE∇↓E_{\nabla} \downarrowEpk↓E_{\mathrm{pk}} \downarrowEobj↓E_{\mathrm{obj}} \downarrowRobj↑R_{\mathrm{obj}} \uparrow
Historical (sealed; Table 4)7.549.5712.537.238.4185.97
Retrained, seed 427.449.6612.527.198.3785.46
Retrained, seed 3141597.439.7012.537.218.5985.55
Retrained, seed 2718287.369.3512.527.208.4085.83
Range, 3 retrainings0.080.340.010.020.230.37
Range, all 4 runs0.180.340.020.040.230.51

Table 12. Seed and platform sensitivity (cohort A, RETROSPECTIVE; mean relative errors and recall, %).

One observed failure case and the remaining errors

When the near field reaches the lid edge. The local-tolerance interpretation has one material exception, not a failure of the mutually exclusive junction/interior partition. The squared-error interaction remains small there (regional interaction at most 1.2×10−31.2 \times10^{-3}), but capsule-02 has by far the largest share of its near-body band in the junction J\mathcal{J} (1.12%; Table 13), where the reference acts. The only other geometry with any, cohort-B capsule-03, has 0.42% and an EnearE_{\mathrm{near}} change of −0.5%.

GeometryPlacementLid gap (box side)N∩J\mathcal{N} \cap\mathcal{J} (% of N\mathcal{N})EvolE_{\mathrm{vol}} (%)EnearE_{\mathrm{near}} (%)EobjE_{\mathrm{obj}} (%)RobjR_{\mathrm{obj}} (pp)
capsule-00Interior0.2420.00−63.3−0.5−0.6+0.04
capsule-01Near z=1z=10.3280.00−65.1−0.3−0.4+0.02
capsule-02Near x/yx/y face0.0781.12−62.2−47.4+4.6−0.01
capsule-03Near z=0z=00.5160.00−62.8+0.2+0.1+0.01
pair-00Interior0.4920.00−62.7+0.1−0.2+0.01
pair-01Near z=1z=10.4840.00−65.30.0−0.2+0.04
pair-02Near x/yx/y face0.2500.00−60.6−0.1+0.1−0.06
pair-03Near z=0z=00.5390.00−65.50.0−0.1+0.02
triplet-00Interior0.3670.00−63.1−0.10.0−0.04
triplet-01Near z=1z=10.2730.00−63.2−0.1−0.2−0.06
triplet-02Near x/yx/y face0.3050.00−63.1+0.1−0.1+0.05
triplet-03Near z=0z=00.2660.00−64.70.00.0+0.04

Table 13. Per-geometry effect of adding the reference (cohort A, means over five Reynolds numbers). “Lid gap” is the solid-to-lid distance in units of the box side; N∩J\mathcal{N} \cap\mathcal{J} is the share of the near-body band inside the lid-edge junction. Placement strata refer to the generator’s target faces; z=1z=1 is not the lid.

  • Its near-body error falls by 47% when the reference is added. The gain comes entirely from the junction cells: its near-body error outside the wall shell changes by +0.01%, and 92% of its lifted near-band error lies in the shell.

  • At the same time its object-peak error rises by 4.6%, the largest local harm in either cohort.

  • Proximity to the lid alone does not produce this effect. Cohort B contains a capsule as close to the lid (gap 0.070 vs. 0.078), with 10.4% of its band inside the lid shell but none in the junction; its EnearE_{\mathrm{near}} changes by +0.3%.

Our hypothesis is that the reference affects body metrics only where the body’s near field reaches materially into the singular corner region. One case supports it and one post hoc comparison is consistent with it, so it is not a characterization.

Object-local tails at high Reynolds number. The composition does not solve the local pressure structure around bodies. Figure 6 shows the case-level object-peak error and recall in both cohorts.

Figure 6. Object-local accuracy per case of our predictor (a, cohort A; b, cohort B; both prospective absolute scores): object-peak error against peak-region recall. Colour encodes the Reynolds number (darker is higher) and marker shape the family. Low recall concentrates at Re⁡≥300\operatorname{Re}\ge300, where the full-volume error is lowest (Table 6).

  • At Re⁡=900\operatorname{Re}=900 mean recall drops to 69.9% (A) and 66.3% (B), from over 90% at Re⁡=10\operatorname{Re}=10, even though the global error is lowest at Re⁡≥300\operatorname{Re}\ge300.

  • The minimum recall, 41.5% (A) and 43.6% (B), occurs in multi-body cases.

  • The worst object-peak errors are 22.8% (A) and 19.7% (B).

Figure 7 shows the worst object-peak and lowest-recall cases next to three reference cases, with object-attached halos and displaced lobes. On 120 training cases, the specified coarse and native target-informed oracles attain near-body errors of 2.65% and 2.49%, and object-peak errors of 3.19% and 2.97%, respectively (Section H). This comparison does not quantify the fraction of representable corrections retained at 643643. In a separate development-data diagnostic, the tested residual learners did not reliably reproduce the target-informed gain; this does not identify a unique learning or representation bottleneck. § 17 gives per-Reynolds, per-family and tail statistics, all case scores and the best cohort-A cases.

Figure 7. Fields on native 1283128^3 planes (cohort A; pressure in units of U2U^2). Rows a–e: a plane crossing both perpendicular lid edges; the only low-Reynolds-number example; capsule-02, whose near-body band reaches into the lid-edge junction; the worst object-peak case; the lowest-recall case. The first four cases are those of the original visual protocol. In xyxy and yzyz planes the lid is the top edge. Columns 1–3 share a symmetric scale set to the 99th percentile of ∣p∣|p| outside the wall shell, so the lid corners saturate by design. Column 4 is the pointwise error of our predictor as a percentage of the case’s full-fluid truth RMS, whose RMS over the fluid equals 100 EvolE_{\mathrm{vol}}; the case RMS printed below converts it to U2U^2. Its scale, 0–20%, is the same in every row and in Figure 10, so rows can be compared directly. Column 5 shows ∣p^corrected−p^lifted∣|\hat p_{\mathrm{corrected}}-\hat p_{\mathrm{lifted}}| computed from the stored, unclipped fields, corresponding to the magnitude of ΠFδB128\Pi_F\delta B_{128} up to storage precision. Its fixed logarithmic scale spans 10−410^{-4}–10−1 U210^{-1}\,U^2; finite fluid values below the lower limit, including exact zeros, use the floor colour. Values above the upper limit saturate. The correction reaches 0.3U20.3U^2 at the perpendicular lid edges (top corners of rows a and e); row b has small but nonzero edge corrections, up to 0.0042U20.0042U^2. Grey is solid in pressure and correction panels and white in error panels; black contours mark the zero level of the signed distance. Small-scale colour variations are visible near the voxelized solid–fluid interfaces.

Numerical robustness: grid, lid and iterative convergence

Numerical robustness: grid, lid and iterative convergence

Three questions remain after § 5.2. Is the junction dominance a property of the singular lid, how does it depend on the grid, and does it depend on how tightly the test truth was converged? We answered them with new CFD under protocols frozen before any run (Sections F and G; retrospective, descriptive). The regularized lid uses ux=[1−(2x−1)18]2[1−(2z−1)18]2u_x = [1 - (2x - 1)^{18}]^2[1 - (2z - 1)^{18}]^2 [23, 7], which vanishes at the lid edges. Everything else is unchanged.

  • Grid refinement. We solved the empty cavity at Re=10,100,997Re = 10, 100, 997 on 64364^3, 1283128^3 and 2563256^3 grids for both lids. The 2563256^3 runs were initialized from the converged 1283128^3 solution, and all runs met the residual criteria.

  • Regularized cohort A. We re-solved all 60 cohort-A cases at 1283128^3 with the regularized lid. We also refitted the reference, by the unchanged procedure, to 17 regularized empty-cavity solutions. All 77 runs converged.

Table 14 gives the grid dependence.

  • Unregularized lid. The mean pressure magnitude in the lid-edge cells, relative to the field RMS, increases at each refinement: 24, 44 and 82 at Re=10Re = 10, and 17, 25 and 41 at Re=997Re = 997. The low-Reynolds sequence nearly doubles; the high-Reynolds ratios are 1.47 and 1.64. These finite-grid observations are consistent with a singular edge response, but do not establish an asymptotic rate for the normalized pressure at high Reynolds number.

  • Energy shares. The share of a band of fixed physical width grows under refinement (66.6%, 72.5%, 76.6% at Re=10Re = 10), as the logarithmic divergence of the energy predicts. The share of a band of fixed cell count falls (83.9%, 72.5%, 62.7%). Neither is resolution independent, so the 69% of § 5.2 is a 1283128^3 number.

  • The residual δB\delta B. It decreases only slowly with the unregularized lid, from 5.5% to 3.9% of the energy at Re=10Re = 10.

  • Regularized lid. The edge pressure appears to stabilize over the tested grids (12.8, 15.0, 16.0 at $Re = 10); the band of physical width 1/16 holds 49–61% of the energy. The residual energy ratio decreases (0.36%, 0.06%, 0.01%). This evidence concerns the specified regularization profile and measured grids, not a proof of the continuum limit for every regularized lid.

With the regularized lid (Table 15), the junction still holds 51.5% of the truth energy, against 68.6% with the unregularized lid. The lid-edge gradients remain steep even without a singularity. For the frozen sharp-lid-trained surrogate, the reference correction does not help (EvolE_{\mathrm{vol}} 31.4% →\to 32.6%). Both the target boundary condition and the reference change in this diagnostic; it does not isolate the effect for a matched regularized-lid-trained model. A smaller δB128\delta B_{128} limits its possible effect but does not determine its sign. The surrogate, trained on unregularized-lid data, now errs mostly at the lid edges: 93.5% of its squared error lies in J\mathcal{J}.

LidArmsJs_JEvol↓E_{\mathrm{vol}}\downarrowEnear↓E_{\mathrm{near}}\downarrowE∇↓E_{\nabla}\downarrowEpk↓E_{\mathrm{pk}}\downarrowEobj↓E_{\mathrm{obj}}\downarrowRobj↑R_{\mathrm{obj}}\uparrowEJE_JEIE_IeJe_J
sharplifted68.620.610.153.521.88.486.024.46.995.1
+ reference7.59.612.57.28.486.07.86.975.4
regularizedlifted51.531.412.7126.534.69.885.543.39.193.5
+ reference32.612.8131.836.09.885.545.19.293.9

Table 15. Regularized lid, cohort A (60 cases at 1283128^3; cohort-mean relative errors and recall, %; retrospective). sJs_J: truth energy share of the junction. “+ reference” adds ΠFδB128\Pi_F \delta B_{128} from a reference refitted to regularized empty-cavity solutions. eJe_J: the junction’s share of the arm’s squared error. The surrogate is unchanged and was trained on unregularized-lid data.

The truth-energy concentration motivates regional reporting under both tested lid profiles. The 93.5% error share additionally reflects a boundary-condition distribution shift, and cannot establish a model-independent regularized-lid error mechanism. The measured benefit of δB128\delta B_{128} is specific to the tested sharp-lid composition and grid pair.

More tightly converged test truth. A predeclared experiment re-solved 12 cohort-A cases, six geometries at Re=10\mathrm{Re}=10 and 900, starting from the terminal state of the test truth (Section G, Table 17). The first arm, A+, keeps the discretization and tightens the linear-solver tolerance to 10−1110^{-11} and the residual controls to 10−810^{-8} (pressure) and 10−910^{-9} (velocity). The second, B, additionally replaces the convection scheme by limitedLinear. For all 12 cases, A+ converged after 3,460–7,245 additional iterations and then changed by at most 3.6×10−83.6 \times10^{-8} over 100 further iterations. It differs from the test truth by at most 1.7×10−51.7 \times10^{-5} globally and 3.8×10−53.8 \times10^{-5} near the bodies. No cohort mean of any metric of ours, the lifted surrogate or the no-learning control moves by more than a relative 2.0×10−52.0 \times10^{-5}, and every predeclared iterative-sensitivity guard passes. B converged in none of these cases within 16,000 additional iterations: its pressure residual stalled at 1.9×10−71.9 \times10^{-7}–7.0×10−77.0 \times10^{-7} at Re=10\mathrm{Re}=10 and at 2.0×10−32.0 \times10^{-3}–3.3×10−33.3 \times10^{-3} at Re=900\mathrm{Re}=900. By the predeclared rule the scheme comparison is therefore unresolved. Tightening test-truth convergence leaves the reported metric effects essentially unchanged, but sensitivity to the convection scheme could not be evaluated from converged alternate-scheme fields.

Table 17: Numerical-reference study (all 12 cases; descriptive). “A+ iterations”: additional SIMPLE iterations to strict convergence. “A+ vs. A”: relative L2L_2 difference of the centered pressure on the full fluid and on the near-body band. “Drift”: the same measure over 100 further iterations. Last column: initial pressure residual of B at its iteration cap, where the strict control is 10−810^{-8}.

GeometryRe\mathrm{Re}A+ iterationsA+ vs. A (10−510^{-5}) globalA+ vs. A (10−510^{-5}) nearA+ drift (10−810^{-8}) globalA+ drift (10−810^{-8}) nearB: pp residual at 16,000
capsule-00107,0960.320.650.180.361.9×10−71.9 \times10^{-7}
capsule-009003,8791.171.631.753.503.3×10−33.3 \times10^{-3}
capsule-01107,2150.561.450.280.675.5×10−75.5 \times10^{-7}
capsule-019003,8561.672.291.702.932.9×10−32.9 \times10^{-3}
pair-00107,1170.251.070.140.627.0×10−77.0 \times10^{-7}
pair-009003,7851.452.211.873.372.0×10−32.0 \times10^{-3}
pair-02107,1470.230.410.130.203.1×10−73.1 \times10^{-7}
pair-029003,6821.121.652.513.042.6×10−32.6 \times10^{-3}
triplet-00107,2451.013.840.491.814.5×10−74.5 \times10^{-7}
triplet-009003,6641.441.171.831.873.2×10−33.2 \times10^{-3}
triplet-03106,8980.502.180.241.071.9×10−71.9 \times10^{-7}
triplet-039003,4601.442.181.883.632.5×10−32.5 \times10^{-3}

Table 17.

Cost

With 48 CPU threads the full predictor needs a median 0.76 s per case after a warm-up call. The surrogate takes 0.45 s, including feature and surface-graph construction, and lift, reference and centering take 0.32 s. Thirty-six replays reproduce the sealed arrays exactly. The native reference CFD took a median 12.7 h (cohort A) and 17.0 h (cohort B) of single-thread wall time per case, with 7,205–13,367 SIMPLE iterations. The method’s offline costs are the 1,620 training solves at 64364^3, the 17 empty-cavity solves at 1283128^3 and training. These numbers are indicative and do not support a matched-resource speed-up claim.

Discussion

Implications for benchmarking. Equation (4) holds for any model, but distinguishes truth-energy weights from model-dependent error shares. In our cavity the junction truth-energy share ranges from 66.6% to 69.8% across the tested Reynolds numbers at 1283128^3 (overall mean 68.6%). For the tested transferred coarse model, the global error primarily reflects its lid-edge error; this is not a universal consequence of the truth weights. Two models can differ by a factor of 2.75 in global error and not at all in the interior. A body-free field can match a body-aware predictor globally, and an improvement near bodies may barely move the global number. The share grows with refinement for a band of fixed physical width and is still about one half with a regularized lid (§ 5.7). We therefore recommend that cavity pressure benchmarks:

  1. report the truth’s junction energy share sJs_J and each model’s error share wJw_J at each Reynolds number and resolution, specifying the band width in both cells and physical units; when comparing grids, report a fixed physical width as well as a fixed cell count;

  • (ii) when sJs_J is large, report errors in the junction, the rest of the wall shell and the interior separately, and near-body errors outside the shell;

  • (iii) report near-body errors also relative to the body-induced field ΠF[p−B]\Pi_F[p-B], where an empty-cavity solution BB is available;

  • (iv) state whether and how the lid is regularized.

The partition requires only masks and regional squared-error sums, which our analysis scripts compute.

What the composition is, and is not. The reference correction reuses offline CFD to supply a body-independent, resolution-specific component that is lost when a coarse field is transferred to the native grid. It is useful when:

  • a learned model must run on a finer grid than it was trained on;

  • the boundary singularity is shared across queries;

  • a small library of body-free solutions is affordable.

It is not a super-resolution method: the learned field is transferred, not refined, and nothing new is learned near the bodies. It is not the best use of the reference either: rebasing the surrogate on it does better on every global metric (§ 5.3). It is also not portable by construction. Its target is a solver- and grid-specific boundary field (Table 14); the corresponding restriction–prolongation residual decreases over the measured regularized-lid refinements.

LidReRe8 cells: 64364^38 cells: 1283128^38 cells: 2563256^3width 1/16: 64364^3width 1/16: 1283128^3width 1/16: 2563256^3δB\delta B: 64364^3δB\delta B: 1283128^3δB\delta B: 2563256^3
sharp1083.972.562.766.672.576.65.484.513.85
sharp10085.171.458.666.471.474.94.713.793.24
sharp99773.165.044.863.065.066.12.721.110.68
regularized1074.148.924.247.248.949.40.360.060.01
regularized10078.051.424.150.051.451.80.300.040.01
regularized99773.060.830.959.860.860.90.430.030.00

Table 14. Grid dependence of the empty cavity (RETROSPECTIVE; shares of the centered pressure energy, %). “8 cells”: share of the cells within eight cells of a lid edge at each resolution. “Width 1/16”: share of the band of that physical width, the same region at every resolution and equal to the 8-cell band at 1283128^3. “δB\delta B”: energy of the restriction–prolongation residual δB=(I−LR2)B\delta B = (I - LR_2)B relative to that of BB.

Limitations.

  • One configuration. One cavity, lid speed, solver and grid pair, tested only at the five training Reynolds numbers.

  • Numerical references. Grid dependence was measured on the empty cavity, not on the test cohorts, whose 2563256^3 solutions would each cost weeks of CPU time. At the lid edges the unregularized fields depend on the grid by nature. Tighter test-truth convergence has negligible metric effects on the 12 tested cases, but the only alternative convection scheme we tried did not converge, so scheme sensitivity remains unresolved.

  • Few independent units. Twelve geometries per cohort, from one procedural generator and three families. Only one test geometry had a material share of its near-body band in the lid-edge junction.

  • Controls on inspected data. The no-learning control, the matched FNO, the seed replication, the diagnostic baselines and the regularized-lid analysis were run after the truth had been inspected. The regularized-lid comparison uses a surrogate trained on unregularized-lid data. The FNO is one untuned configuration, and no native-1283128^3-trained baseline was run; the 95% junction error share is specific to transferred coarse models.

  • Open object-local problem. Object-local errors at high Reynolds number remain unsolved. Our volumetric near-body metrics are not surface pressure or force, and the AI-agent visual reviews were not blinded or a substitute for human validation.

Conclusion

For the tested 643→128364^3 \to128^3 transferred surrogate and sharp-lid OpenFOAM reference fields, most of the pressure squared norm and surrogate error sits in 1.5% of the fluid at the lid edges, predominantly at the two edges where the lid scrapes against a side wall. Adding a body-free, Reynolds-conditioned reference to the frozen surrogate lowers mean full-volume relative L2L_2 by 63.5% (20.6% to 7.5%) in the prospective comparison on 12 unseen geometries. A post hoc reconstructed contrast on 12 more shows a similar global reduction. In the matched control, the learned term carries the substantial interior and near-body improvement. A body-free field alone scores about as well globally as the body-aware predictor, and rebasing on the reference lowers the global error further with little change in cohort-mean body metrics. Separating the two makes the evaluation interpretable. The junction truth-energy share depends on the lid profile and band definition: it grows under refinement at fixed physical width, decreases at fixed cell count, and remains substantial for the tested regularized profile. The reported metric effects are insensitive to tighter test-truth convergence in the 12 tested cases. The headline improvement is a lid-edge correction, and the unsolved problem is the pressure structure around bodies at high Reynolds number.

Broader Impact Statement

This work concerns surrogate models for laminar incompressible flow in a canonical configuration. We are not aware of direct negative societal impacts. Our main practical message is cautionary: a global error metric can overstate how well a surrogate captures the physics near objects. Surrogate predictions should not replace validated simulation in engineering decisions without region-resolved checks.

Funding and Competing Interests

This work was funded by EXDYNA Pte. Ltd. Several authors are affiliated with EXDYNA Pte. Ltd., as listed on the first page. The authors declare no other competing interests.

Reproducibility Statement

The supplementary material contains:

  • the geometry recipes of the 12 cohort-A test geometries;

  • a cohort-A score-replay package containing four-arm metric means, four-arm per-case volume errors, the two prospectively compared arms’ per-case metrics, and signed regional squared-error savings for the reference and learned-term contrasts. A complementary accounting file provides all 540 case–region records, including cell counts, truth and four-arm error squared norms, body-response denominator energies and source identities; its script reconstructs regional means and separately labeled arithmetic/geometric-ratio statistics.

  • for the diagnostic baselines and the numerical-robustness study, the per-case result records, the protocols frozen before the runs, the solver and analysis scripts and a script that regenerates the corresponding tables;

  • manifests linking the packaged records and replay outputs to their source identities and file hashes, together with dated original prediction-seal and truth-attempt metadata. These are author-maintained protocol records, not independently certified timestamps; hashes alone do not establish temporal order.

The replay packages verify the included scores, spatial accounting and numerical-study tables, and rebuild the score-only effect plot (Figure 3); they do not regenerate field figures or reproduce field inference, training or CFD solutions. Model weights, CFD arrays and training code are not included in the submission. Sections B to D, F and G describe the architecture, training, reference fit, geometry construction, solver settings and the new runs.

Metric, transfer and decision-rule details

Centering and supports. Prediction and truth are centered separately on the same mask m=1[SDF>0]m = 1[\mathrm{SDF} > 0] and zeroed outside the fluid, so a constant shift of either raw field leaves every metric unchanged. The near-body band N\mathcal{N} is open at SDF =0= 0 and closed at SDF =0.2= 0.2 (unit-cube units). The object-local region further excludes three cells at each outer face; this margin is defined in cells, so its physical thickness differs between 64364^{3} and 1283128^{3}.

Gradient error. For axis aa, let VaV_a contain cell centers whose center and both axis neighbours are fluid, excluding the two end planes of that axis. With daqi=(qi+ea−qi−ea)/2d_aq_i = (q_{i+e_a} - q_{i-e_a})/2,

E∇=[∑a∑i∈Va(dap^i−dapi)2∑a∑i∈Va(dapi)2]1/2(10)E_{\nabla} = \left[\frac{\sum_a\sum_{i\in V_a}(d_a\hat{p}_i-d_ap_i)^2}{\sum_a\sum_{i\in V_a}(d_ap_i)^2}\right]^{1/2} \tag*{(10)}

This pools energy across axes. It is a relative discrepancy of discrete gradients, not an equation residual. It includes the wall shell: only one stencil cell is eroded.

High-magnitude regions and recall. A region with N≥20N \ge20 eligible cells selects k=N−⌊0.95N⌋k = N - \lfloor0.95N\rfloor cells by stable descending sort of ∣p∣|p| in C-order (ties resolved by the lower flattened index). A flat prediction selects an empty set and has zero recall. A flat truth region, with range ≤10−12max⁡(1,max⁡∣p∣)\le10^{-12}\max(1,\max|p|), is invalid. Relative-field norms ≤10−15\le10^{-15} and gradient squared norms ≤10−24\le10^{-24} are rejected. RobjR_{\mathrm{obj}} measures overlap of absolute-pressure supports, not local-maximum detection or signed-lobe agreement; EobjE_{\mathrm{obj}} measures the signed field error on the truth-selected cells.

Transfer. For a source cell-centered field qq with source fluid support msm_s and target support mtm_t,

Lm(q)j={I(msq)j/I(ms)j,mt,j=1 and I(ms)j>10−6,0,otherwise,(11)L_m(q)_j = \begin{cases} I(m_sq)_j/I(m_s)_j, & m_{t,j}=1\text{ and }I(m_s)_j > 10^{-6},\\ 0, & \text{otherwise}, \end{cases} \tag*{(11)}

where II is trilinear interpolation at source-index coordinates (j+12)ns/nt−12(j+\frac{1}{2})n_s/n_t-\frac{1}{2} with nearest extension at the domain boundary. A constant source value is reproduced on supported cells. The transfer is rejected when fewer than ⌊0.99∣Ft∣⌋\lfloor0.99|\mathcal{F}_t|\rfloor target cells are supported. Accumulation is float64 and output float32. R2R_2 is 232^3 block averaging, so Equation (5) is a restriction–prolongation residual, not a spectral projection; R2δB=0R_2\delta B = 0 is not assumed.

Percentiles. p90 and p10 use linear interpolation at zero-based indices 0.9(n−1)0.9(n-1) and 0.1(n−1)0.1(n-1) of the sorted case scores (NumPy’s default).

Predeclared decision rule for cohort A

Before the cohort-A truth was generated we fixed the following criteria.

  • Absolute thresholds on the reference-corrected predictor. Maximum mean/p90 errors: .12/.18 for EvolE_{\mathrm{vol}}, .25/.40 for EnearE_{\mathrm{near}}, .18/.28 for E∇E_{\nabla}, .12/.20 for EpkE_{\mathrm{pk}} and .25/.40 for EobjE_{\mathrm{obj}}. Minimum recall mean/p10: .65/.35. All were met (Table 16).

  • Relative to the lifted arm, cohort level. At least 10% mean reduction in EvolE_{\mathrm{vol}}, E∇E_{\nabla} and EpkE_{\mathrm{pk}}. At most 2% mean increase in EnearE_{\mathrm{near}} and EobjE_{\mathrm{obj}}. At most .01 absolute loss in RobjR_{\mathrm{obj}}.

  • Relative to the lifted arm, per geometry, family and Reynolds group. A joint condition: lower mean EvolE_{\mathrm{vol}}, at most 2% higher EnearE_{\mathrm{near}} and EobjE_{\mathrm{obj}}, and at most .01 lower RobjR_{\mathrm{obj}}. Required coverage: at least 9/12 geometries, 2/3 families and 4/5 Reynolds groups.

  • Visual check: shared-scale inspection for new artifacts such as frames, checkerboards, halos, displaced geometry or amplitude collapse.

CohortStatisticEvol↓E_{\mathrm{vol}}\downarrowEnear↓E_{\mathrm{near}}\downarrowE∇↓E_{\nabla}\downarrowEpk↓E_{\mathrm{pk}}\downarrowEobj↓E_{\mathrm{obj}}\downarrowRobj↑R_{\mathrm{obj}}\uparrow
AMean7.59.612.57.28.486.0
Ap90 (recall: p10)8.412.314.98.513.567.1
AWorst9.220.215.18.622.841.5
BMean7.49.812.57.39.783.3
Bp90 (recall: p10)8.313.714.98.516.265.3
BWorst8.515.215.28.619.743.6

Table 16. Distribution of the reference-corrected predictor’s scores (60 cases per cohort, %; prospective absolute scores). For recall the tail is the 10th percentile and the worst is the minimum.

The joint condition held in 11/12 geometries, 3/3 families and 5/5 Reynolds groups, and no new artifact was found. These are application-specific performance criteria, not significance tests.

Coarse surrogate: specification

Coarse surrogate: specification

Inputs and background. At unit-cube cell center xx the six dense channels are:

  • SDF/3\sqrt{3};

  • the three components of 2x−12x-1;

  • twice the nearest outer-wall distance;

  • r∗=2(log⁡10Re⁡−1)/(log⁡10900−1)−1r_* = 2(\log_{10}\operatorname{Re}-1)/(\log_{10}900-1)-1, which maps Re⁡=10\operatorname{Re}=10 to −1-1 and Re⁡=900\operatorname{Re}=900 to +1+1.

All channels are zeroed where SDF ≤0\le0. For family ff and Reynolds number rr, with 54 training cases Tf,rT_{f,r}, fluid masks mkm_k and centered pressures pkcp_k^c,

bf,r(i)=∑k∈Tf,rmk(i)pkc(i)∑k∈Tf,rmk(i),b64(r)=16∑f=16bf,r,(12)b_{f,r}(i)=\frac{\sum_{k\in T_{f,r}}m_k(i)p_k^c(i)}{\sum_{k\in T_{f,r}}m_k(i)},\qquad b_{64}(r)=\frac{1}{6}\sum_{f=1}^{6}b_{f,r}, \tag*{(12)}

with zero at unsupported cells. No family label is used at inference.

Objective. With pc=ΠF64pp^c=\Pi_{F_{64}}p and p^c=ΠF64(b64+cθ)\hat{p}^c=\Pi_{F_{64}}(b_{64}+c_\theta), the loss is

L=LF+LN+Lpk+0.1L∇,LA=∑i∈A(p^ic−pic)2∑i∈A(pic)2.(13)\mathcal{L}=\mathcal{L}_{F}+\mathcal{L}_{N}+\mathcal{L}_{\mathrm{pk}}+0.1\mathcal{L}_{\nabla},\qquad\mathcal{L}_{A}=\frac{\sum_{i\in A}(\hat{p}_i^c-p_i^c)^2}{\sum_{i\in A}(p_i^c)^2}. \tag*{(13)}

The peak support is ∣pc∣|p^c| at or above the 95th percentile within N\mathcal{N}, after excluding three outer-wall cells. L∇\mathcal{L}_{\nabla} is the pooled gradient ratio defining E∇2E^2_{\nabla}. Losses compare complete centered fields, not the neural residual alone.

Optimization. Training uses the following settings:

  • 1,620 cases from 324 geometries, with 40,500 single-case updates;

  • AdamW with learning rate 3×10−43\times10^{-4}, weight decay 10−510^{-5} and per-update cosine decay;

  • bf16 autocast, gradient clipping at 1.0 and activation checkpointing;

  • seed 42 for the sealed run;

  • the terminal checkpoint fixed in advance, with no validation-based selection and no early stopping.

Volume pathway. A 131^3 input projection, one residual block, four restriction stages (encoder widths 24, 32, 48, 64, 96), two bottleneck blocks and four upsampling stages with skip connections. Each residual block has two 333^3 support-normalized convolutions,

[(K∗(sv))⋅27max⁡(13∗s,1)+b]1[s>0],(14)\left[(K*(s v))\cdot\frac{27}{\max(1^3*s,1)}+b\right]\mathbf{1}[s>0], \tag*{(14)}

each preceded by channelwise LayerNorm, with SiLU and a residual multiplier initialized at 0.1. Restriction returns sv/max⁡(sˉ,10−6)s v/\max(\bar{s},10^{-6}). A zero-initialized 131^3 head gives cθc_\theta after fluid masking and centering.

Surface pathway. Marching cubes on the sampled 64364^3 SDF gives triangles, grouped by connected solid component and nearest fluid owner cell into area-weighted surface nodes with 13 features (position, normal, log-area, wall distance, offset from the component centroid, component area fraction, r∗r_*). Graph neighbours are the 12 nearest nodes of the same component. At each of nine exchanges (five encoder, four decoder), two message-passing layers update a width-48 surface state. The layers use the edge features [hi,hj−hi,δij,∥δij∥,ni⋅nj,δij⋅ni,δij⋅nj][h_i,h_j-h_i,\delta_{ij},\lVert\delta_{ij}\rVert,n_i\cdot n_j,\delta_{ij}\cdot n_i,\delta_{ij}\cdot n_j]. The state then gathers volume features at owner cells. In the reverse direction, node states are area-averaged into one global token. The token is fused into active cells within unit-cube distance 0.25 of the surface, with weight (1−d/0.25)2(1-d/0.25)^2. Figure 8 shows the architecture. The network has 3,150,088 parameters.

Figure 8. Coarse surrogate architecture (grid and channel labels are exact; cube sizes are schematic). a, Multiscale volume pathway: a 3D U-Net-style encoder (light blue) and decoder (darker blue) joined by matched skips (thin lines) (Ronneberger et al., [39]; Çiçek et al., [11]). Brown marks the surface pathway, including the nine bidirectional volume–surface exchanges (E). b, The four surface-coupling components that differ from the U-Net backbone; the numbers only index the boxes. The learned correction cθc_{\theta} is added to the Reynolds-only background b64(Re)b_{64}(\mathrm{Re}) and centered on the 64364^{3} fluid by ΠF64\Pi_{F_{64}}.

Reference fit and fidelity

Each of the 17 empty-cavity pressures, at Re⁡=10\operatorname{Re}=10, 13, 18, 24, 32, 42, 56, 75, 100, 133, 177, 237, 316, 421, 561, 748 and 997, is centered, and the ensemble mean μ\mu is subtracted. For X=UΣVTX=U\Sigma V^{\mathsf{T}} the first eight rows of VTV^{\mathsf{T}} give orthonormal modes ϕj\phi_j and UkjΣjjU_{kj}\Sigma_{jj} the training coefficients. Reconstruction is

B128(Re⁡)=Πcavity[μ+∑j=18aj(log⁡10Re⁡) ϕj](15)B_{128}(\operatorname{Re})=\Pi_{\mathrm{cavity}}\left[\mu+\sum_{j=1}^{8}a_j(\log_{10}\operatorname{Re})\,\phi_j\right] \tag*{(15)}

with shape-preserving piecewise-cubic interpolants aja_j and no extrapolation. Mean and basis are stored in float32 and coefficients are accumulated in float64. Rank eight was fixed before any test outcome. The fit was validated before this study on 16 held-out native-grid empty-cavity solutions at Re⁡=11,15,…,864\operatorname{Re}=11,15,\ldots,864 (Table 3). Visual inspection of the three worst cases (Re⁡=864,365,648\operatorname{Re}=864,365,648) found no artifacts. Of the five test Reynolds numbers, 30, 300 and 900 are interpolated and 10 and 100 are fit points. At the 17 fit points the rank-8 reconstruction differs from the library fields by at most 3.5×10−43.5\times10^{-4} in relative L2L_2 norm (Re⁡=561\operatorname{Re}=561), so fit points are reproduced closely but not exactly.

Geometry construction, CFD settings and cost

Geometry. Frames are sampled from normalized Gaussian directions with Gram–Schmidt orthogonalization.

  • Capsules join a cylinder of length U(0.18,0.28)U(0.18,0.28) and radius U(0.055,0.075)U(0.055,0.075) to same-radius end spheres.

  • Pairs combine a sphere of radius U(0.060,0.085)U(0.060,0.085) with an oriented box whose half-sizes lie between (0.045,0.050,0.055)(0.045,0.050,0.055) and (0.070,0.080,0.090)(0.070,0.080,0.090). The center separation is the sphere radius plus the box bounding radius plus U(0.025,0.055)U(0.025,0.055).

  • Triplets place a sphere (U(0.050,0.065)U(0.050,0.065)), a cylinder (radius U(0.035,0.045)U(0.035,0.045), length U(0.10,0.14)U(0.10,0.14)) and a box (half-sizes from (0.035,0.040,0.045)(0.035,0.040,0.045) to (0.050,0.055,0.060)(0.050,0.055,0.060)) at −s-s, 0, +s+s along a frame axis.

Placement strata constrain the bounding-sphere gap to the selected face to [0.025,min⁡(0.07,0.48−rb)][0.025,\min(0.07,0.48-r_b)]. Generator seeds are 2026090901 (cohort A) and 2026091407 (cohort B). All 24 geometries are disjoint by hash from all training and development geometries. Supplementary Data S1 lists the 12 accepted cohort-A recipes. SDFs are unions of exact primitive SDFs without smoothing. The box SDF is ϕ(∣RT(x−c)∣−h)\phi(\lvert R^{\mathsf{T}}(x-c)\rvert-h) with ϕ(v)=∥v+∥2+min⁡(max⁡jvj,0)\phi(v)=\lVert v_+\rVert_2+\min(\max_j v_j,0), and cylinders have flat caps.

Solver settings. All 120 test cases were rendered from one solver-control template (SHA-256 prefix 4177e7a6). The settings are:

  • steadyState; Gauss linear gradients; bounded Gauss linearUpwind grad(U) convection; Gauss linear corrected Laplacians;

  • GAMG (pressure) and smoothSolver (velocity) with Gauss–Seidel smoothing, tolerance 10−910^{-9} and relative tolerance 0.01;

  • SIMPLE with no non-orthogonal correctors, relaxation 0.3/0.5 and residual controls 10−6/10−710^{-6}/10^{-7};

  • zero initial fields and zero-gradient pressure on all walls; the lid velocity is not smoothed at the edges.

Acceptance required reported convergence within 16,000 iterations, normal termination and finite continuity diagnostics. After stripping comments, fvSchemes and fvSolution are byte-identical between the test cases and all 33 native empty-cavity cases, i.e. the 17 library cases and the 16 held-out cases of Table 3. The library’s controlDict differs only in restarting from the latest time after an iteration-ceiling extension and in writing six rather than twelve significant digits. We re-verified the 64364^3 cases retrospectively: of all 2,874 64364^3 cases on the compute host, 2,863 have fvSchemes and fvSolution identical to the test template after comment stripping. The other 11 are solver diagnostics from two development experiments, not training cases; they differ only in residual tolerances tightened to 10−910^{-9} and 10−1010^{-10}. Terminal global continuity errors of the cohort-A truth are at most 2.7×10−182.7 \times10^{-18} in magnitude. Failed-job and retry histories are incomplete; every prespecified test case is reported.

Cost. Single-thread simpleFoam wall time per test case was 6.9–16.3 h (median 12.7 h) in cohort A and 9.5–20.3 h (median 17.0 h) in cohort B, with 48 and 60 concurrent single-thread jobs, respectively. Cases took 7,205–13,367 iterations. Surrogate timing is given in § 5.8.

Least-squares representation residual and diagnostic baselines

Least-squares representation residual. For a centered field qq on the 1283128^3 fluid mask, the target optimization is

η∗(q)=min⁡u∈R∣F64∣, c∈R∥q−Lmu−c1∥2,F∥q∥2,F.\eta_*(q) = \min_{u\in\mathbb{R}^{|F_{64}|},\,c\in\mathbb{R}} \frac{\lVert q-L_m u-c\mathbf{1}\rVert_{2,F}}{\lVert q\rVert_{2,F}}.

We report residuals attained by LSQR using the sparse matrix of the mask-normalized transfer LmL_m (Section A), which reproduces the transfer of the paper to 1.2×10−71.2 \times10^{-7}. LSQR uses atol = btol = 10−1010^{-10} and at most 4,000 iterations. Convergence diagnostics were not retained, so these estimates are not certified optima. The attained residuals upper-bound η∗(q)\eta_*(q); they do not certify a lower bound on achievable error and apply only to this transfer, not to all 64364^3 representations or decoders.

For the cohort-A truth the numerical residual is 13.8% on average (cohort B: 13.8%). Evaluating the same globally fitted residual within each region gives case-mean relative norms of 16.6% in J\mathcal{J}, 14.8% in the wall shell, 3.3% on the object-peak support, 2.9% in the interior near-body band and 1.0% in the interior. These are not separately optimized regional lower bounds. The case-mean share of residual energy in J\mathcal{J} is 99.9%; this is not the fraction of truth energy lost. For the empty-cavity reference (Re=10,30,100,300,900\mathrm{Re}=10, 30, 100, 300, 900), the attained residuals are 17.1%, 17.0%, 15.7%, 11.4% and 8.4%, against 21.2%, 21.1%, 19.5%, 14.2% and 10.6% for ∥δB128∥/∥B128∥\lVert\delta B_{128}\rVert/\lVert B_{128}\rVert. These smaller feasible residuals show that block averaging followed by the lift is not optimal; its residual norms are about a quarter larger than the attained LSQR residuals, without certifying the unrepresentable part.

Voxelization. Between the 64364^3 and 1283128^3 masks of the same geometry, the combined count of fine-fluid cells with a solid parent and fine-solid cells with a fluid parent is at most 0.064% of the fine-fluid count. On cohort B the fine mask regenerated from the geometry recipe equals the truth mask in every case.

Diagnostic baselines. The rebased surrogate of Equation (8) uses 64364^3 empty-cavity solutions at the five test Reynolds numbers. They were computed with the reference library’s case files, the mesh changed to 64364^3, and all converged within 3,643 iterations. In cohort B the lifted arm is the reconstruction ΠF[p^128−ΠFδB128]\Pi_F[\hat p_{128}-\Pi_F\delta B_{128}]. Applied to five cohort-A cases, where the lifted arm was scaled, this reconstruction reproduces the scaled arrays to a relative error of at most 4.2×10−84.2 \times10^{-8}. For e=p^−pe=\hat p-p and the set EF\mathcal{E}_F of interior faces joining two fluid cells, the face error is the relative norm

Eface=[∑(i,j)∈EF(ei−ej)2∑(i,j)∈EF(pi−pj)2]1/2.E_{\mathrm{face}}=\left[\frac{\sum_{(i,j)\in\mathcal{E}_F}(e_i-e_j)^2}{\sum_{(i,j)\in\mathcal{E}_F}(p_i-p_j)^2}\right]^{1/2}.

The restriction–transfer residual energy ratio is ∥e−IR2e∥22/∥e∥22\lVert e-IR_2e\rVert_2^2/\lVert e\rVert_2^2, evaluated on the full grid with solids zero-filled, where R2R_2 is 2×2×22 \times2 \times2 block averaging and II is ordinary trilinear interpolation with nearest boundary extension, without mask normalization. Since IR2IR_2 is not an orthogonal projection, this ratio is neither a frequency-band energy fraction nor a checkerboard-only fraction.

Grid and regularized-lid study: protocol and runs

The protocol was frozen and committed before any run. Every case copies an existing, validated case (mesh, schemes, linear solvers and residual controls) and changes only the lid profile or the mesh resolution; the regularized profile is evaluated at the lid face centres. Iteration ceilings were 30,000 (64364^3, 1283128^3) and 40,000 (2563256^3); convergence status is reported and no run was dropped or repeated after its result was seen.

  • Regularized cohort A: 60 cases, 7,511–13,645 SIMPLE iterations, all converged.

  • Regularized reference library: 17 empty-cavity cases at the library Reynolds numbers, 7,640–12,713 iterations, all converged; refitted with rank 8 and the unchanged interpolation.

  • Grid study: empty cavity at Re=10\mathrm{Re}=10, 100, 997, both lids. 64364^3: 2,580–3,592 iterations. 2563256^3: eight MPI ranks, initialized by consistent field mapping from the converged 1283128^3 solution of the same lid and Reynolds number, then 4,285–11,522 iterations to the unchanged residual criteria. Solving 2563256^3 from rest was measured at about 22 s per iteration, which is why the initialization was changed before any 2563256^3 result existed.

The regional analysis of the regularized cohort uses the unchanged predictor, lift, centering, metric functions and regions. A self-check that applies the same code to the unregularized cohort reproduces the frozen reference fit to 1.9×10−61.9 \times10^{-6} and the sealed predictions to 3.8×10−63.8 \times10^{-6}.

Numerical-reference study: protocol and runs

Cases, arms, thresholds and the decision precedence were frozen before any run. The 12 cases are capsules 00 and 01, pairs 00 and 02 and triplets 00 and 03 (in each family the interior placement and one other placement stratum), each at Re=10\mathrm{Re}=10 and 900. Both arms start from independent copies of the terminal UU, pp and ϕ\phi of the test truth A on the same native mesh and may use at most 16,000 additional SIMPLE iterations.

  • A+A^+: the discretization of A; linear-solver tolerance 10−1110^{-11} with zero relative tolerance; residual controls 10−810^{-8} (pressure) and 10−910^{-9} (velocity).

  • BB: as A+A^+, with bounded Gauss limitedLinear 1 convection instead of linearUpwind.

After its strict endpoint each arm runs exactly 100 further iterations without residual control. The relative change of the centered pressure over these iterations is the drift, with validity limits 10−410^{-4} globally and 2×10−42 \times10^{-4} near the bodies. An unconverged arm makes the comparison unresolved, and no later rule can override this. The runs used CPU jobs on Google Cloud with the pinned OpenFOAM v2412 runtime.

In the original jobs the drift endpoint was not written. After a restart OpenFOAM’s time index is not aligned with the time value, and with a write interval of 100 the step strict+100 was skipped. The strict endpoints, meshes and dictionaries had been stored, so we repeated the drift invocation from the stored strict endpoints with a write interval of 1 and otherwise identical dictionaries. For one case we compared the repeated solver log with the original: all 100 iterations have identical residuals, and only the cumulative continuity diagnostic, which is not restored on restart, differs.

Development history and coarse-grid evidence

The composition was chosen on development data, before cohort A existed. Table 18 lists the distinct cohorts. None is pooled with the native-grid cohorts.

Table 18: Evaluation cohorts. Cases are CFD solutions; geometries are independent units. “Prospective” means predictions were fixed before the reference pressure was generated. †For our predictor’s absolute scores; the lifted contrast is post hoc.

CohortRoleCases/geom.GridComparison
Coarse replication 1prospective120/2464364^3volume baseline (text)
Coarse replication 2prospective120/2464364^3volume and spectral baselines (Table 19)
Surface cohortprospective120/2464364^3surface-conditioned vs. volume model
Surface ablationdevelopment60/1264364^3complete vs. null surface input
Composition cohortprospective80/1664364^3coarse surrogate on new compositions
Freeform cohortprospective80/1664364^3coarse surrogate on freeform shapes
Cohort Aprospective60/121283128^3this paper, main contrast
Cohort Bprospective†60/121283128^3this paper, replication

Table 18.

Coarse-grid baselines. Four models share one coarse replication cohort (Table 19). All use 900 training cases, 22,500 updates and seed 42. The background-residual masked volume model (BR-MGL) has 2.0M parameters; the FNO has width 16, four blocks and 636^3 modes (1.77M real parameters). Residual models predict a correction to a family/Reynolds background; direct models predict the full pressure. MGL denotes the masked volume model and BR-MGL its background-residual form. On coarse replication 1, BR-MGL gave Evol=5.28%E_{\mathrm{vol}} = 5.28\%, Enear=13.59%E_{\mathrm{near}} = 13.59\%, E∇=4.07%E_{\nabla} = 4.07\%, Epk=2.65%E_{\mathrm{pk}} = 2.65\%, Eobj=10.45%E_{\mathrm{obj}} = 10.45\% and Robj=80.28%R_{\mathrm{obj}} = 80.28\%. On separate prospective cohorts (Table 20), the surface-conditioned model improves on the volume model. A matched null-surface ablation on development data gives EnearE_{\mathrm{near}} 10.7% vs. 12.1%, EobjE_{\mathrm{obj}} 7.7% vs. 8.8% and RobjR_{\mathrm{obj}} 83.3% vs. 80.0%. The coarse surrogate generalizes to new analytic compositions and to occupancy-derived freeform shapes at 64364^3. Its 64364^3 global error there (3.1–3.7%), measured on its own training grid, is much smaller than the 1283128^3 global error of the reference-corrected predictor (7.4–7.5%; lifted surrogate 20.6%), while its near-body error is similar (9.7–9.9% vs. 9.6–9.8%).

ModelEvol↓E_{\mathrm{vol}} \downarrowEnear↓E_{\mathrm{near}} \downarrowE∇↓E_{\nabla} \downarrowEpk↓E_{\mathrm{pk}} \downarrowEobj↓E_{\mathrm{obj}} \downarrowRobj↑R_{\mathrm{obj}} \uparrow
BR-MGL (residual)5.2613.023.672.139.1880.17
Direct MGL10.1713.0315.488.389.0880.13
Residual FNO6.1314.884.842.3511.3078.34
Direct FNO13.0416.3720.9710.3211.3877.19

Table 19. Coarse-grid volume and spectral baselines (coarse replication 2, 120 cases/24 geometries; relative errors and recall in %). MGL: masked volume model; “BR” and “residual”: correction to a family/Reynolds background; “direct”: full pressure.

CohortPredictorEvol↓E_{\mathrm{vol}} \downarrowEnear↓E_{\mathrm{near}} \downarrowE∇↓E_{\nabla} \downarrowEpk↓E_{\mathrm{pk}} \downarrowEobj↓E_{\mathrm{obj}} \downarrowRobj↑R_{\mathrm{obj}} \uparrow
SurfaceBR-MGL5.1813.003.792.439.3180.29
Surface-conditioned4.129.813.381.836.6883.95
CompositionBackground only7.1139.064.842.8041.0654.23
Coarse surrogate3.119.682.631.247.3585.94
FreeformBackground only9.0147.595.523.7848.7449.62
Coarse surrogate3.679.882.911.527.5384.54

Table 20. Separate coarse-grid prospective cohorts (relative errors and recall in %). Horizontal rules separate datasets. In the surface cohort, the surface-conditioned model is an earlier variant with a family/Reynolds background; in the other two cohorts, “coarse surrogate” is the frozen surrogate of this paper and “background only” its Reynolds-only background b64b_{64}.

Lifting versus direct execution. On 12 development cases (six geometries, Re=10,900\mathrm{Re} = 10, 900), the lifted surrogate had Evol/Enear/E∇/Eobj=19.9/11.4/49.9/9.9%E_{\mathrm{vol}}/E_{\mathrm{near}}/E_{\nabla}/E_{\mathrm{obj}} = 19.9/11.4/49.9/9.9\%. Executing the same weights directly at 1283128^3 gave 20.6/30.7/50.0/25.7%, and adding the reference to the lift gave 8.4/11.4/12.0/9.9%. Direct execution changes the physical scale of all learned operators and produced wider body shells and displaced lobes.

Development energy diagnostic. On the same development cases, the coarse-to-fine CFD discrepancy ΠF[p128c−Lmp64c]\Pi_F[p_{128}^{c}-L_m p_{64}^{c}] had a case-mean 99.3% of its energy in the eight-cell wall shell outside the near-body band. The reference residual δB128\delta B_{128} had more than 99.9%. This motivated the composition; the post hoc regional analysis of § 5.2 replaces it.

Headroom and learnability of the remaining near-body error. Two target-informed oracle diagnostics illustrate possible near-body corrections on development and training data. Both use truth and are not predictors or optimality bounds.

  • Learnability. On the same 12 development cases, an oracle restricted the centered truth-minus-prediction residual of our composition to 64364^3, recentered it, lifted it back with the transfer of Section A, added it and recentered again. This reduced the remaining error. Residual learners using only deployment-visible inputs, with fresh or transferred initialization, did not reliably reproduce this gain. Under an equal 3,000-update budget, training on 24 geometries instead of a repeated set of 12 improved EobjE_{\mathrm{obj}} but still worsened EnearE_{\mathrm{near}} by 6.08% and lowered RobjR_{\mathrm{obj}} by 0.037 relative to the fixed composition.

  • Resolution. On 120 training cases from 24 geometries at all five Reynolds numbers, both oracles add a tapered, centered truth residual r=ΠF(p−pb)r=\Pi_F(p-p_b) to the fixed composition pbp_b: ΠF[pb+w T(wr)]\Pi_F[p_b+w\,T(wr)]. The weight ww is 1 for 0<SDF≤0.20<SDF\leq0.2, 0 outside (0,0.35)(0,0.35), and 1−3u2+2u31-3u^2+2u^3 in between, with u=(SDF−0.2)/0.15u=(SDF-0.2)/0.15. For the coarse oracle, TT is support-normalized restriction to 64364^3 followed by lifting. For the native oracle, TT is fluid-mask-normalized Gaussian smoothing at 1283128^3 (σ=1\sigma=1 cell). The coarse oracle reaches Enear=2.65%E_{\mathrm{near}}=2.65\%, Eobj=3.19%E_{\mathrm{obj}}=3.19\% and Robj=98.8%R_{\mathrm{obj}}=98.8\%; the native oracle reaches 2.49%, 2.97% and 98.8%.

On the same data, the native-resolution oracle improves on the 64364^3 oracle by 5.9% (EnearE_{\mathrm{near}}) and 7.1% (EobjE_{\mathrm{obj}}). These two constructions attain similar near-body errors, but do not measure the fraction of all representable corrections retained at 64364^3. The first diagnostic shows that the tested learners did not reliably reproduce the target-informed gain. Representability does not establish learnability, and these unequal-degree-of-freedom constructions do not bound all representations.

Additional results

Geometry-level statistics and families

Table 21 gives the post hoc geometry-level statistics behind Table 4: medians, leave-one-geometry-out ranges, one-sided bounds and sign-flip tests. Table 22 gives the prospective absolute scores of our predictor by family in both cohorts; object-local errors are lowest for capsules in both.

CohortFamilyEvol↓E_{\mathrm{vol}}\downarrowEnear↓E_{\mathrm{near}}\downarrowE∇↓E_{\nabla}\downarrowEpk↓E_{\mathrm{pk}}\downarrowEobj↓E_{\mathrm{obj}}\downarrowRobj↑R_{\mathrm{obj}}\uparrow
Acapsule7.68.112.57.24.887.7
Apair7.610.712.77.210.287.8
Atriplet7.59.912.57.310.382.4
Bcapsule7.48.512.47.37.481.1
Bpair7.510.812.57.311.883.7
Btriplet7.310.012.57.210.185.0

Table 22. Families (our predictor, prospective absolute scores; cohort means in %).

The sign-flip calculation enumerates all 2122^{12} sign patterns of the paired geometry-mean differences. Inferential exactness requires joint sign symmetry under the null (for example, independent symmetric paired differences), not geometry independence alone. The tt bounds assume approximately normal geometry-level log ratios for errors or paired differences for recall. These are unadjusted, metric-wise descriptive summaries for the sampled generator and placement strata, not simultaneous guarantees or tests of the predeclared arithmetic cohort-mean thresholds.

All case scores and linked diagnostics

Figure 9 shows every case score of our predictor in both cohorts. It includes two shape diagnostics: the correlation ρ\rho between predicted and true centered fields, and the amplitude ratio aa, i.e. predicted over true standard deviation. Their cohort-A means are ρ=0.998\rho= 0.998 and a=0.954a = 0.954. For each case Evol2=1+a2−2aρE_{\mathrm{vol}}^2 = 1 + a^2 - 2a\rho, so these are linked to EvolE_{\mathrm{vol}}, not independent evidence. High correlation and nearly correct amplitude coexist with the object-local errors of § 5.5.

Figure 9. All 60 case scores per cohort for our predictor (prospective absolute scores). Filled circles, cohort A; open squares, cohort B; black bars, cohort means. Points are jittered across their lane for display only. a, Relative errors. b, Peak-region recall. c, One minus the field correlation ρ\rho. d, Amplitude ratio aa, predicted over true field standard deviation; the dashed line marks a=1a = 1.

Best cases

For balance with the failure cases of Figure 7, Figure 10 shows the best cohort-A cases under a fixed rule:

Figure 10. Best cohort-A cases under the rule stated above: lowest EvolE_{\mathrm{vol}} (also lowest EobjE_{\mathrm{obj}}), lowest EnearE_{\mathrm{near}}, highest RobjR_{\mathrm{obj}}, and lowest multi-body EvolE_{\mathrm{vol}}; each row shows the xyxy plane with the most solid cells. Columns 1–2 share a symmetric scale set to the 99th percentile of ∣p∣\lvert p\rvert outside the wall shell, as in Figure 7. Column 3 is the pointwise error of our predictor as a percentage of the case’s full-fluid truth RMS, with the case RMS printed below to convert to U2U^2, on the same 0–20% scale as Figure 7. Columns 4–5 enlarge the dashed box (all bodies plus an 8-cell margin) on the box’s own symmetric scale, set to the 99.5th percentile of ∣p∣\lvert p\rvert in the box. Grey is solid in pressure panels and white in error panels; black contours mark the zero level of the signed distance. Short dark marks on body outlines are first-layer fluid cells of the voxelized boundary, where the CFD pressure itself has local extremes.

  • the lowest EvolE_{\mathrm{vol}}, which is also the lowest EobjE_{\mathrm{obj}} (capsule-01, Re=300\mathrm{Re} = 300: 5.8% and 2.6%);

  • the lowest EnearE_{\mathrm{near}} (capsule-01, Re=30\mathrm{Re} = 30: 5.8%);

  • the highest RobjR_{\mathrm{obj}} (pair-02, Re=100\mathrm{Re} = 100: 97.5%);

  • the lowest EvolE_{\mathrm{vol}} among multi-body cases (triplet-03, Re=300\mathrm{Re} = 300: 5.8%).

Shown is the xyxy plane with the most solid cells. The selection depends on the scores, so these cases illustrate what the method does well; they do not represent the cohort.

Gradient and peak metrics by region

Table 8 repeats the regional analysis of § 5.2 for E∇E_{\nabla} and EpkE_{\mathrm{pk}}. It was computed by a separately frozen regional audit of the same arrays; the audit’s additive identity was checked at the documented float32 tolerance, with maximum deviation 7.3×10−67.3 \times10^{-6}. The audit also decomposes squared errors into regional offset and within-region shape. Interior offsets account for about 1% of interior squared error in both arms, so the interior result is not an artifact of global centering. That audit’s frozen descriptive rule labels the reference gain “spatially mixed” rather than “boundary dominated”. The only failing input is the pooled near-body squared-error change by Reynolds number (−22%-22\% to −61%-61\%), which capsule-02’s junction cells dominate. Every other input is below 10−410^{-4} against a limit of 0.05, and excluding capsule-02 the near-body changes are within ±0.5%\pm0.5\%. We report the frozen label unchanged.

Learned-increment accounting

Relative to the no-learning control, the squared error that our predictor saves is split over the 12 wall-by-body regions. 64.8% of it lies in object-local fluid and 76.2% outside the wall shell when each case is normalized by its truth energy; pooled over cases these shares are 61.2% and 69.5%. Arrays reconstructed for this audit match the sealed ones to 7.3×10−67.3 \times10^{-6} (pressure units), and replayed metrics match to 8.3×10−168.3 \times10^{-16}.

Seed replication

After the method and the cohort-A result were frozen, the complete surrogate training was repeated for the predeclared seeds 42, 314159 and 271828 in a cloud environment; their evaluations on the consumed cohort A are retrospective. Architecture, data, ordering rule, objective, optimizer, budget, composition and evaluator were unchanged; only initialization, RNG and ordering seed varied. Every seed passes the predeclared absolute and learned-increment checks. The retrained seed-42 run differs from the historical seed-42 run by 0.10 points in EvolE_{\mathrm{vol}} and 0.51 points in RobjR_{\mathrm{obj}}, so run-to-run variation includes platform effects. Table 12 therefore reports the four-run envelope.

Matched FNO: decision rule

The comparison rule was fixed before any FNO training and is applied in both directions, per equal-seed pair. These are retrospective comparisons on cohort A. The rule has nine numerical checks and a visual veto:

  • the winner improves mean EnearE_{\mathrm{near}} and EobjE_{\mathrm{obj}} by at least 5% each;

  • it raises mean RobjR_{\mathrm{obj}} by at least 0.02;

  • it regresses mean EvolE_{\mathrm{vol}}, E∇E_{\nabla} and EpkE_{\mathrm{pk}} by at most 2% each;

  • it regresses the p90 of EnearE_{\mathrm{near}} and EobjE_{\mathrm{obj}} by at most 2% each;

  • it wins jointly on EnearE_{\mathrm{near}} and EobjE_{\mathrm{obj}} in at least 9/12 geometries, with at most 0.01 recall loss.

Per seed, the reference-corrected surrogate lowers the FNO’s EnearE_{\mathrm{near}} by 21.8–25.8% and EobjE_{\mathrm{obj}} by 22.7–27.6%, raises recall by 3.2–5.1 points, and wins jointly in 10–11 of 12 geometries. Inspection of 24 fixed-scale panels shows broader smooth residuals and stronger interface lobes for the FNO, and no artifact that reverses the numerical result. An earlier, unmatched direct FNO had Evol=13.5%E_{\mathrm{vol}} = 13.5\% and Enear=17.0%E_{\mathrm{near}} = 17.0\% under the same composition; it used fewer training cases and updates.

Origin of cohort B

Cohort B was generated to test a later candidate: an equal-weight ensemble of a different surface-conditioned model, composed with the same reference. The reference-corrected predictor of this paper served as that test’s sealed control. That candidate improved EobjE_{\mathrm{obj}} by 8.7% but did not meet its own promotion rule, and it is not part of this paper. Here cohort B supplies prospective absolute scores of our predictor, a post hoc reconstructed reference contrast and regional and least-squares residual diagnostics, and retrospective reference-alone and rebased-surrogate baselines. The new geometries do not make these later-designed analyses prospective.

References

  1. [1]S. Albensoeder and H. C. Kuhlmann. Accurate three-dimensional lid-driven cavity flow. Journal of Computational Physics, 206(2):536–558, 2005.DOI
  2. [2]B. Alkin, M. Bleeker, R. Kurle, et al. AB-UPT: Scaling neural CFD surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers. Transactions on Machine Learning Research, 2025.arxiv.org/abs/2502.09692
  3. [3]F. Bartolucci, E. de Bézenac, B. Raonic, R. Molinaro, S. Mishra, and R. Alaifari. Representation equivalent neural operators: a framework for alias-free operator learning. In Advances in Neural Information Processing Systems, 2023.arxiv.org/abs/2305.19913
  4. [4]J. Berner, M. Liu-Schiaffini, J. Kossaifi, et al. Principled approaches for extending neural architectures to function spaces for operator learning. Nature Machine Intelligence, 8:1173–1181, 2026.arxiv.org/abs/2506.10973
  5. [5]F. Bonnet, J. A. Mazari, P. Cinnella, and P. Gallinari. AirfRANS: High fidelity computational fluid dynamics dataset for approximating Reynolds-Averaged Navier–Stokes solutions. In Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, 2022.arxiv.org/abs/2212.07564
  6. [6]O. Botella and R. Peyret. Benchmark spectral results on the lid-driven cavity flow. Computers & Fluids, 27 (4):421–433, 1998.DOI
  7. [7]R. Bouffanais, M. O. Deville, and E. Leriche. Large-eddy simulation of the flow in a lid-driven cubical cavity. Physics of Fluids, 19:055108, 2007.
  8. [8]W. X. Chan, W. Ding, B. Li, H. S. Wong, and C. H. Yap. Role of physics-informed constraints in real-time estimation of 3D vascular fluid dynamics using multi-case neural network. Computers in Biology and Medicine, 190:110074, 2025.
  9. [9]W. Chen, Z.-F. Wei, Y. Fu, M. Penwarden, P. Roy, and P. Stinis. ArGEnT: Arbitrary geometry-encoded transformer for operator learning. arXiv preprint arXiv:2602.11626, 2026.arxiv.org/abs/2602.11626
  10. [10]Z. Cheng, Z. Hao, X. Wang, et al. Reference neural operators: Learning the smooth dependence of solutions of PDEs on geometric deformations. arXiv preprint arXiv:2405.17509, 2024.
  11. [11]Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger. 3D U-Net: Learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2016.arxiv.org/abs/1606.06650
  12. [12]F. de Avila Belbute-Peres, T. D. Economon, and J. Z. Kolter. Combining differentiable PDE solvers and graph neural networks for fluid flow prediction. In Proceedings of the 37th International Conference on Machine Learning, pp. 2402–2411, 2020.
  13. [13]M. Elrefaie, F. Morar, A. Dai, and F. Ahmed. DrivAerNet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks. In Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, 2024.
  14. [14]M. S. Eshaghi, Y. Wang, N. Valizadeh, X. Zhuang, and T. Rabczuk. Neural operators for immersed-boundary soft swimmers locomotion. arXiv preprint arXiv:2608.07722, 2026.arxiv.org/abs/2608.07722
  15. [16]C. Hancock, E. Lewis, and H. K. Moffatt. Effects of inertia in forced corner flows. Journal of Fluid Mechanics, 112:315–327, 1981.DOI
  16. [17]B. N. Hanna, N. T. Dinh, R. W. Youngblood, and I. A. Bolotnov. Machine-learning based error prediction approach for coarse-grid Computational Fluid Dynamics (CG-CFD). Progress in Nuclear Energy, 118:103140, 2020.
  17. [18]Z. Hao, J. Yao, C. Su, et al. PINNacle: A comprehensive benchmark of physics-informed neural networks for solving PDEs. In Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, 2024.
  18. [19]P. Heger, D. Hilger, M. Full, and N. Hosters. Investigation of physics-informed deep learning for the prediction of parametric, three-dimensional flow based on boundary data. Computers & Fluids, 278:106302, 2024.
  19. [21]H. C. Kuhlmann and F. Romanò. The lid-driven cavity. In Computational Modelling of Bifurcations and Instabilities in Fluid Dynamics, pp. 233–309. Springer, 2019.DOI
  20. [22]Anna Kwa, Spencer K. Clark, Brian Henn, Noah D. Brenowitz, Jeremy McGibbon, Oliver Watt-Meyer, W. Andre Perkins, Lucas Harris, and Christopher S. Bretherton. Machine-learned climate model corrections from a global storm-resolving model: Performance across the annual cycle. Journal of Advances in Modeling Earth Systems, 15:e2022MS003400, 2023. doi: 10.1029/2022MS003400. URL https://doi.org/10.1029/2022MS003400.
  21. [23]E. Leriche and S. Gavrilakis. Direct numerical simulation of the flow in a lid-driven cubical cavity. Physics of Fluids, 12(6):1363–1376, 2000.
  22. [27]N. Liu, S. Jafarzadeh, and Y. Yu. Domain agnostic Fourier neural operators. arXiv preprint arXiv:2305.00478, 2023.arxiv.org/abs/2305.00478
  23. [29]Y. Luo, Y. Chen, and Z. Zhang. CFDBench: A large-scale benchmark for machine learning methods in fluid dynamics. arXiv preprint arXiv:2310.05963, 2024.arxiv.org/abs/2310.05963
  24. [30]N. McGreivy and A. Hakim. Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations. Nature Machine Intelligence, 6:1256–1269, 2024.DOI
  25. [31]H. K. Moffatt. Viscous and resistive eddies near a sharp corner. Journal of Fluid Mechanics, 18(1):1–18, 1964.
  26. [32]B. A. Nosek, C. R. Ebersole, A. C. DeHaven, and D. T. Mellor. The preregistration revolution. Proceedings of the National Academy of Sciences, 115(11):2600–2606, 2018.
  27. [33]L. Oakden-Rayner, J. Dunnmon, G. Carneiro, and C. Ré. Hidden stratification causes clinically meaningful failures in machine learning for medical imaging. In Proceedings of the ACM Conference on Health, Inference, and Learning, pp. 151–159, 2020.
  28. [34]R. Ohana, M. McCabe, L. Meyer, et al. The well: a large-scale collection of diverse physics simulations for machine learning. In Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, 2024.arxiv.org/abs/2412.00568
  29. [35]E. Pajaziti, J. Montalt-Tordera, C. Capelli, et al. Shape-driven deep neural networks for fast acquisition of aortic 3D pressure and velocity flow fields. PLOS Computational Biology, 19(4):e1011055, 2023.DOI
  30. [36]A. Rabeh, E. Herron, A. Balu, et al. Benchmarking scientific machine-learning approaches for flow prediction around complex geometries. Communications Engineering, 4:182, 2025a.DOI
  31. [37]A. Rabeh, A. Krishnamurthy, and B. Ganapathysubramanian. 3D neural operator-based flow surrogates around 3D geometries: Signed distance functions and derivative constraints. arXiv preprint arXiv:2503.17289, 2025b.arxiv.org/abs/2503.17289
  32. [38]R. Ranade, M. A. Nabian, K. Tangsali, et al. DoMINO: A decomposable multi-scale iterative neural operator for modeling large scale engineering simulations. arXiv preprint arXiv:2501.13350, 2025.arxiv.org/abs/2501.13350
  33. [39]O. Ronneberger, P. Fischer, and T. Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015.arxiv.org/abs/1505.04597
  34. [40]Ehsan Roohi and Christopher J. McDevitt. Regional accuracy of physics-informed neural networks (PINNs) for recirculation in a deep lid-driven cavity. arXiv:2610.38814v2, 2026. URL https://arxiv.org/abs/2610.38814v2.
  35. [42]A. Shaa, C. Guet, X. Yang, A. Albergel, and B. Ribstein. Hybrid neural interpolation of a sequence of wind flows. Engineering Applications of Computational Fluid Mechanics, 20(1):2682648, 2026.DOI
  36. [43]P. N. Shankar and M. D. Deshpande. Fluid mechanics in the driven cavity. Annual Review of Fluid Mechanics, 32:93–136, 2000.DOI
  37. [44]P. Sharpe. GLOBE: Accurate and generalizable PDE surrogates using domain-inspired architectures and equivariances. arXiv preprint arXiv:2511.15856, 2025.
  38. [45]M. Takamoto, T. Praditia, R. Leiteritz, D. MacKinlay, F. Alesiani, D. Pflüger, and M. Niepert. PDEBench: An extensive benchmark for scientific machine learning. In Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, 2022.arxiv.org/abs/2210.07182
  39. [46]R. Tali, A. Rabeh, C.-H. Yang, et al. FlowBench: A large scale benchmark for flow simulation over complex geometries. arXiv preprint arXiv:2409.18032, 2024.arxiv.org/abs/2409.18032
  40. [47]G. I. Taylor. On scraping viscous fluid from a plane surface. In M. Schäfer (ed.), Miszellan­een der Angewandten Mechanik, pp. 313–315. Akademie-Verlag, Berlin, 1962.
  41. [48]N. Thuerey, K. Weissenow, L. Prantl, and X. Hu. Deep learning methods for Reynolds-averaged Navier–Stokes simulations of airfoil flows. AIAA Journal, 58(1):25–36, 2020.DOI
  42. [50]H. Wang, J. Li, A. Dwivedi, K. Hara, and T. Wu. BENO: Boundary-embedded neural operators for elliptic PDEs. In International Conference on Learning Representations, 2024.arxiv.org/abs/2401.09323
  43. [51]Q. Wang, C. Wang, M. Zhang, J. Sun, P. Yang, S. Tang, and S. Xiang. MNO: Multiscale neural operator for 3D computational fluid dynamics. arXiv preprint arXiv:2510.16071, 2026.arxiv.org/abs/2510.16071
  44. [52]L. Wei, Y. Zhao, Z. Du, Q. Xu, and F. Zhou. Blade3DNO: Geometry-consistent spectral operator learning for transonic 3D compressor flows. Aerospace Science and Technology, 177:112334, 2026.DOI
  45. [53]H. G. Weller, G. Tabor, H. Jasak, and C. Fureby. A tensorial approach to computational continuum mechanics using object-oriented techniques. Computers in Physics, 12(6):620–631, 1998.DOI
  46. [54]S. Wen, A. Kumbhat, L. Lingsch, S. Mousavi, Y. Zhao, P. Chandrasekar, and S. Mishra. Geometry aware operator transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains. In Advances in Neural Information Processing Systems, 2025.arxiv.org/abs/2505.18781
  47. [55]X. Zhang, Y. Huang, S. Jiang, Z. Wang, and M. Jiang. Geometry-aware anisotropic boundary correction for aerodynamic simulation. arXiv preprint arXiv:2606.09963, 2026.arxiv.org/abs/2606.09963
  48. [56]J. Zhou, V. Duruisseaux, D. Z. Huang, and A. Anandkumar. Boundary-augmented neural operators for better generalization to unseen geometries. In NeurIPS Workshop on AI for Science, 2025.

Paper details

vlnc:W_q4jmuk3w
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

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