Organizations: School of Computing, National University of Singapore, Singapore · EXDYNA Pte. Ltd., 81 Ayer Rajah Crescent, #02-49, Singapore 139967 · Tandon School of Engineering, New York University, Brooklyn, NY, USA · Institute of Advanced Intelligence and Computing (IAIC), Agency for Science, Technology, and Research (A*STAR), Singapore, 138632
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 1283 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 643 surrogate lifted to 1283 lowers the mean full-volume relative L2 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.
Figures & tables
Figure 1: Where the error lives (cohort A, 60 native 1283 cases). a , Centered CFD pressure magnitude relative to the case’s fluid RMS on the xy plane z=43.5/128 of cohort-A heterogeneous triplet-03 at Re=300 (log scale; grey, solid). The lid moves in +x . White boxes mark the junction cells 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∣ 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 Section 3.3 .
Method
Solution information available
Learned vs. deterministic operations
Reference Neural Operators ( Cheng et al., 2024 )
reference-domain solution, reference and query geometry, a constructed deformation
transport the reference; learn the query-minus-transported correction
Multifidelity DeepONet ( Lu et al., 2022 )
low- and high-fidelity training data; low-fidelity predictions as inputs
learn a low-fidelity operator and a high-fidelity residual
Hybrid wind interpolation ( Shaa et al., 2026 )
CFD snapshots for a fixed object and sampled inlet angles
interpolation ansatz inside a trained neural model
Ours
immersed-geometry training fields; an offline body-free library; query geometry and Re only at inference
freeze the surrogate; add the deterministic residual (I−LR2)B128 ; center
Table 1: Information used by reference-informed approaches. These are methodological distinctions, not numerical comparisons.
Metric
Definition
Evol
Equation 2 on the full fluid volume F
Enear
Equation 2 on the near-body band N
E∇
Relative error of axiswise central differences where the center and both neighbours are fluid
Epk
Equation 2 on the truth’s top-5% ∣p∣ cells of F (global peaks)
Eobj
Equation 2 on the truth’s top-5% ∣p∣ cells of N , excluding three cells at the outer walls (object peaks)
Robj
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 Appendix A ). All are dimensionless. The near-body band N={0<SDF≤0.2} is 25.6 cells thick at 1283 and overlaps the wall shell for most bodies; only Eobj and Robj exclude cells at the outer walls.
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=0 inside p^64 , keeping the coarse background b64 . No query pressure and no query-time CFD enter the prediction.
Held-out Re (16 values, 11–864)
Mean
p90
Worst
Re 11–37 / 48–205 / 274–864
Evol of B128(Re) (%)
0.072
0.164
0.198
0.039 / 0.051 / 0.131
E∇ of B128(Re) (%)
0.071
0.162
0.244
–
1−ρ , field correlation ( 10−6 )
0.37
1.34
1.87
–
Table 3: Held-out-Reynolds fidelity of the empty-cavity reference (obstacle-free; 16 held-out values). Relative errors in %; 1−ρ in units of 10−6 ( ρ : 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).
Evol↓
Enear↓
E∇↓
Epk↓
Eobj↓
Robj↑
Cohort A (primary): 60 cases, 12 unseen geometries
Lifted surrogate
20.6
10.1
53.5
21.8
8.4
86.0
Reference-corrected (ours)
7.5
9.6
12.5
7.2
8.4
86.0
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 improved
12/12
7/12
12/12
12/12
8/12
8/12
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% t 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.
Arm
Evol↓
Enear↓
E∇↓
Epk↓
Eobj↓
Robj↑
EJ
EI
Eface
Cohort A
Lifted surrogate
20.6
10.1
53.5
21.8
8.4
86.0
24.4
6.9
73.1
Reference-corrected (ours)
7.5
9.6
12.5
7.2
8.4
86.0
7.8
6.9
10.1
Reference alone, ΠFB128
8.9
43.9
3.6
3.3
39.3
55.8
1.4
20.5
3.3
Rebased on the native reference
3.2
9.4
2.5
1.2
8.2
86.0
0.7
6.7
2.6
Cohort B
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) . EJ and EI are the junction and interior errors of Table 7 ; Eface is the relative error of differences across interior faces between fluid cells.
Evol (%)
Ours (%)
Cohort
Re
Lifted
Ours
Change
Enear
Eobj
Robj
Worst Eobj
Worst Robj
A
10
24.8
8.2
− 66.8
8.8
8.6
92.2
17.1
86.8
A
30
24.6
8.2
− 66.9
8.7
8.2
91.9
14.1
85.7
A
100
22.8
7.6
− 66.7
9.6
8.9
88.0
15.9
75.0
A
300
17.4
6.9
− 60.5
10.7
7.4
87.8
13.5
81.0
A
900
13.5
6.8
− 49.5
10.0
9.0
69.9
22.8
41.5
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 Evol when adding the reference; “worst” is the largest error or smallest recall among the 12 cases.
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).
Share (%)
Regional error ER (%)
Change of ER (%)
Truth
Cohort A
Cohort B
Region R
Cells
energy
Lifted
No-learning
Ours
+Reference
+Learned
+Reference
Junction J
1.5
68.6
24.4
7.9
7.8
− 67.8
− 0.3
− 67.7
Rest of wall shell S∖J
31.7
18.0
5.7
9.1
5.3
− 6.7
− 42.2
–
Interior I
66.9
13.4
6.9
17.3
6.9
+ 0.2
− 59.9
+ 0.2
Near body
12.3
2.1
10.0
40.0
10.0
0.0
− 75.0
–
Table 7: Regional decomposition (cohort A unless marked; post hoc ). Cell and truth-energy shares are case means. ER applies Equation 2 on region R 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 ; “–”, not computed for cohort B.
Truth
Share of sq. error (%)
Regional error (%)
Median ratio
Metric
Region
share (%)
Lifted
Ours
Lifted
Ours
ours/lifted
E∇
Wall shell S
98.7
99.7
–
53.8
12.3
–
Junction J
94.5
99.4
90.1
–
–
–
Interior I
1.3
0.3
5.3
22.6
22.6
–
Epk
Junction J
17.2 ∗
98.9
92.5
–
–
0.32
Interior I
32.1 ∗
0.3
2.2
–
–
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 Epk the J mask is the audit’s lid-edge band, whose per-case errors coincide with those on J .
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 ( Section 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 1 c: open circle, lifted; triangle, no-learning control; filled dot, ours. b , Cell distance to the immersed surface within the interior 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 1.4–4.9 × worse.
Truth energy
Regional error (%)
Subregion
Cells
share (%)
Lifted
No-learning
Ours
ΠFB128 alone
Perpendicular edges ( x=0,1 )
14,336
59.9
24.4
7.7
7.7
1.1
Parallel edges ( z=0,1 )
14,336
0.8
5.7
7.5
5.4
8.1
Vertex blocks
2,048
7.9
25.4
8.9
8.9
1.2
Junction J (all)
30,720
68.6
24.4
7.9
7.8
1.4
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 83 vertex blocks. The last column evaluates the body-free reference ΠFB128(Re) alone against the truth.
Width k
Cells
Truth energy
Lifted sq. error
Change inside
Change outside
(cells)
(%)
share (%)
share (%)
Jk (%)
Jk (%)
1
0.02
18.5
88.4
− 85.0
+ 6.8
2
0.10
31.5
91.3
− 75.3
+ 3.9
4
0.38
49.3
93.2
− 69.7
− 9.0
8
1.47
68.6
95.1
− 67.8
− 3.3
12
3.20
77.9
95.7
− 67.3
− 0.8
Table 10: Sensitivity to the junction width (cohort A, post hoc ). Jk contains the cells within k cells of a lid edge (the bins of Figure 4 a); changes are those of the cohort-mean regional error when adding the reference. “Outside” includes the rest of the wall shell; for k≤2 it contains the cell layers in which the reference over-corrects slightly.
Arm (cohort A)
Status
Evol↓
Enear↓
E∇↓
Epk↓
Eobj↓
Robj↑
Background only
r
21.8
38.3
53.6
22.1
36.4
61.9
Lifted surrogate
p
20.6
10.1
53.5
21.8
8.4
86.0
No-learning control
r
10.1
38.0
12.9
7.9
36.4
61.8
Reference-corrected (ours)
p
7.5
9.6
12.5
7.2
8.4
86.0
Ours, 3 retrained seeds (mean)
r
7.4
9.6
12.5
7.2
8.5
85.6
Matched residual FNO, 3 seeds (mean)
r
7.9
12.6
12.7
7.3
11.4
81.7
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.
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% .
Run (cohort A)
Evol↓
Enear↓
E∇↓
Epk↓
Eobj↓
Robj↑
Historical (sealed; Table 4 )
7.54
9.57
12.53
7.23
8.41
85.97
Retrained, seed 42
7.44
9.66
12.52
7.19
8.37
85.46
Retrained, seed 314159
7.43
9.70
12.53
7.21
8.59
85.55
Retrained, seed 271828
7.36
9.35
12.52
7.20
8.40
85.83
Range, 3 retrainings
0.08
0.34
0.01
0.02
0.23
0.37
Range, all 4 runs
0.18
0.34
0.02
0.04
0.23
0.51
Table 12: Seed and platform sensitivity (cohort A, retrospective ; mean relative errors and recall, %).
Lid gap
N∩J
Change when adding the reference
Geometry
Placement
(box side)
(% of N )
Evol (%)
Enear (%)
Eobj (%)
Robj (pp)
capsule-00
Interior
0.242
0.00
− 63.3
− 0.5
− 0.6
+ 0.04
capsule-01
Near z=1
0.328
0.00
− 65.1
− 0.3
− 0.4
+ 0.02
capsule-02
Near x / y face
0.078
1.12
− 62.2
− 47.4
+ 4.6
− 0.01
capsule-03
Near z=0
0.516
0.00
− 62.8
+ 0.2
+ 0.1
+ 0.01
pair-00
Interior
0.492
0.00
− 62.7
+ 0.1
− 0.2
+ 0.01
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 is the share of the near-body band inside the lid-edge junction. Placement strata refer to the generator’s target faces; z=1 is not the lid.
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 , where the full-volume error is lowest ( Table 6 ).
Figure 7: Fields on native 1283 planes (cohort A; pressure in units of U2 ). 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 xy and yz planes the lid is the top edge. Columns 1–3 share a symmetric scale set to the 99th percentile of ∣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 100Evol ; the case RMS printed below converts it to U2 . 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∣ computed from the stored, unclipped fields, corresponding to the magnitude of ΠFδB128 up to storage precision. Its fixed logarithmic scale spans 10−4 – 10−1U2 ; finite fluid values below the lower limit, including exact zeros, use the floor colour. Values above the upper limit saturate. The correction reaches 0.3U2 at the perpendicular lid edges (top corners of rows a and e); row b has small but nonzero edge corrections, up to 0.0042U2 . 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.
8 cells
width 1/16
δB
Lid
Re
643
1283
2563
643
1283
2563
643
1283
2563
sharp
10
83.9
72.5
62.7
66.6
72.5
76.6
5.48
4.51
3.85
sharp
100
85.1
71.4
58.6
66.4
71.4
74.9
4.71
3.79
3.24
sharp
997
73.1
65.0
44.8
63.0
65.0
66.1
2.72
1.11
0.68
regularized
10
74.1
48.9
24.2
47.2
48.9
49.4
0.36
0.06
0.01
regularized
100
78.0
51.4
24.1
50.0
51.4
51.8
0.30
0.04
0.01
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 1283 . “ δB ”: energy of the restriction–prolongation residual δB=(I−LR2)B relative to that of B .
Lid
Arm
sJ
Evol↓
Enear↓
E∇↓
Epk↓
Eobj↓
Robj↑
EJ
EI
eJ
sharp
lifted
68.6
20.6
10.1
53.5
21.8
8.4
86.0
24.4
6.9
95.1
+ reference
7.5
9.6
12.5
7.2
8.4
86.0
7.8
6.9
75.4
regularized
lifted
51.5
31.4
12.7
126.5
34.6
9.8
85.5
43.3
9.1
93.5
+ reference
32.6
12.8
131.8
36.0
9.8
85.5
45.1
9.2
93.9
Table 15: Regularized lid, cohort A (60 cases at 1283 ; cohort-mean relative errors and recall, %; retrospective ). sJ : truth energy share of the junction. “+ reference” adds ΠFδB128 from a reference refitted to regularized empty-cavity solutions. eJ : the junction’s share of the arm’s squared error. The surrogate is unchanged and was trained on unregularized-lid data.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Cohort
Statistic
Evol↓
Enear↓
E∇↓
Epk↓
Eobj↓
Robj↑
A
Mean
7.5
9.6
12.5
7.2
8.4
86.0
A
p90 (recall: p10)
8.4
12.3
14.9
8.5
13.5
67.1
A
Worst
9.2
20.2
15.1
8.6
22.8
41.5
B
Mean
7.4
9.8
12.5
7.3
9.7
83.3
B
p90 (recall: p10)
8.3
13.7
14.9
8.5
16.2
65.3
B
Worst
8.5
15.2
15.2
8.6
19.7
43.6
Appendix
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.
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., 2015 ; Çiçek et al., 2016 ) . 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 cθ is added to the Reynolds-only background b64(Re) and centered on the 643 fluid by ΠF64 .
A +
A + vs. A ( 10−5 )
A + drift ( 10−8 )
B: p residual
Geometry
Re
iterations
global
near
global
near
at 16,000
capsule-00
10
7,096
0.32
0.65
0.18
0.36
1.9×10−7
capsule-00
900
3,879
1.17
1.63
1.75
3.50
3.3×10−3
capsule-01
10
7,215
0.56
1.45
0.28
0.67
5.5×10−7
capsule-01
900
3,856
1.67
2.29
1.70
2.93
2.9×10−3
pair-00
10
7,117
0.25
1.07
0.14
0.62
7.0×10−7
Appendix
Table 17: Numerical-reference study (all 12 cases; descriptive). “A + iterations”: additional SIMPLE iterations to strict convergence. “A + vs. A”: relative L2 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 10−8 .
Cohort
Role
Cases/geom.
Grid
Comparison
Coarse replication 1
prospective
120/24
643
volume baseline (text)
Coarse replication 2
prospective
120/24
643
volume and spectral baselines ( Table 19 )
Surface cohort
prospective
120/24
643
surface-conditioned vs. volume model
Surface ablation
development
60/12
643
complete vs. null surface input
Composition cohort
prospective
80/16
643
coarse surrogate on new compositions
Freeform cohort
prospective
80/16
643
coarse surrogate on freeform shapes
Appendix
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 .
Model
Evol↓
Enear↓
E∇↓
Epk↓
Eobj↓
Robj↑
BR-MGL (residual)
5.26
13.02
3.67
2.13
9.18
80.17
Direct MGL
10.17
13.03
15.48
8.38
9.08
80.13
Residual FNO
6.13
14.88
4.84
2.35
11.30
78.34
Direct FNO
13.04
16.37
20.97
10.32
11.38
77.19
Appendix
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.
Cohort
Predictor
Evol↓
Enear↓
E∇↓
Epk↓
Eobj↓
Robj↑
Surface
BR-MGL
5.18
13.00
3.79
2.43
9.31
80.29
Surface-conditioned
4.12
9.81
3.38
1.83
6.68
83.95
Composition
Background only
7.11
39.06
4.84
2.80
41.06
54.23
Coarse surrogate
3.11
9.68
2.63
1.24
7.35
85.94
Freeform
Background only
9.01
47.59
5.52
3.78
48.74
49.62
Coarse surrogate
3.67
9.88
2.91
1.52
7.53
84.54
Appendix
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 b64 .
Change of
Median
LOGO
One-sided 95%
Geometries
Sign-flip
Cohort
Metric
means (%)
geometry (%)
range (%)
bound (%)
improved
p
A
Evol
− 63.46
− 63.16
[ − 63.72, − 63.28]
− 62.74
12/12
0.0005
A
Enear
− 5.53
− 0.08
[ − 6.11, − 0.05]
+ 4.28
7/12
0.2168
A
E∇
− 76.59
− 76.70
[ − 76.69, − 76.55]
− 76.39
12/12
0.0005
A
Epk
− 66.89
− 66.93
[ − 66.96, − 66.82]
− 66.66
12/12
0.0005
A
Eobj
+ 0.08
− 0.09
[ − 0.12, + 0.13]
+ 0.95
8/12
0.9854
Appendix
Table 21: Geometry-level statistics for adding the reference ( n=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 ( t on log ratios); † lower 95% bound on the mean paired recall change. Non-inferiority margins are +2% for Enear,Eobj and −1 pp for Robj . In cohort A, the Enear bound exceeds its margin because of the favourable capsule-02 outlier ( Section 5.5 ); excluding it post hoc, the bound is below +0.1% , which does not establish non-inferiority for the full cohort.
Cohort
Family
Evol↓
Enear↓
E∇↓
Epk↓
Eobj↓
Robj↑
A
capsule
7.6
8.1
12.5
7.2
4.8
87.7
A
pair
7.6
10.7
12.7
7.2
10.2
87.8
A
triplet
7.5
9.9
12.5
7.3
10.3
82.4
B
capsule
7.4
8.5
12.4
7.3
7.4
81.1
B
pair
7.5
10.8
12.5
7.3
11.8
83.7
B
triplet
7.3
10.0
12.5
7.2
10.1
85.0
Appendix
Table 22: Families (our predictor, prospective absolute scores; cohort means in %).
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 . d , Amplitude ratio a , predicted over true field standard deviation; the dashed line marks a=1 .
Figure 10: Best cohort-A cases under the rule stated above: lowest Evol (also lowest Eobj ), lowest Enear , highest Robj , and lowest multi-body Evol ; each row shows the xy plane with the most solid cells. Columns 1–2 share a symmetric scale set to the 99th percentile of ∣p∣ 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 U2 , 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∣ 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.
Neural surrogate models for computational fluid dynamics (CFD) are typically trained as forward operators that map explicit problem specifications, such as geometry and boundary conditions, to solution fields. This ties the model to the conditioning variables seen during training and limits reuse under boundary-condition shifts or local geometry changes. We propose to reformulate steady CFD inference as an inpainting problem: instead of training on explicit boundary conditions, we learn a self-supervised prior over velocity fields and impose boundary constraints only during inference by fixing known regions such as inlet, outlet or unchanged regions from previous simulations. To scale this idea to large 3D meshes, we introduce a local neighbourhood tokeniser that represents high-resolution velocity fields as compact spatial latent tokens and train latent flow-matching and masked-autoencoder models on these tokens. On intracranial aneurysm hemodynamics, our method reconstructs full velocity fields from sparse boundary context, outperforms supervised neural surrogates under boundary-condition and dataset shift and enables local geometry editing by reusing unchanged simulation context. These results suggest that viewing CFD inference as context-conditioned inpainting can turn neural surrogates from task-specific predictors into reusable flow priors.
Jonas Weidner, Yeray Martin-Ruisanchez, Daniel Rueckert +2
AI for Image-Guided Diagnosis and Therapy, Technical University of Munich · Munich Center for Machine Learning (MCML) · AI in Healthcare and Medicine, Technical University of Munich +1
Aerodynamic surrogate models trained on high-fidelity CFD data reproduce numerical predictions of both scalar outputs and entire fields accurately, yet their predictive fidelity is limited by systematic discrepancies between CFD and experimental observations. We present an experimentally grounded correction framework that adapts a CFD-trained deep learning surrogate using wind-tunnel PSP measurements. A Geotransolver surrogate trained on 2,300 high-fidelity CFD simulations of the NASA CRM wing-body configuration, spanning geometric variation, Mach 0.70-0.85, and angles of attack 0 to 4 degrees, reproduces the CFD integrated aerodynamic forces and pitching moment to R2 > 0.99 but does not match the experimental data. To incorporate experimental information without retraining the surrogate, a correction network is trained on spatially registered PSP measurements at two freestream Mach numbers (0.70 and 0.85) across the same angle-of-attack range, learning the discrepancy between the surrogate-predicted and experimentally measured surface-pressure distributions. At Mach 0.85 the correction substantially improves agreement with PSP, particularly at the wing suction peak, shock location, and subsequent pressure recovery, reducing both the magnitude of the prediction error and the fraction of wetted surface on which it exceeds 0.05 in Cp, and it does so from a limited experimental dataset without modifying the pretrained surrogate parameters. On held-out angles of attack the grounded surrogate agrees with measurement to within 2.3-2.7% of the measured Cp range, and outperforms direct interpolation between the measured conditions at every state tested. Experimental measurements can therefore ground a large-scale simulation-trained surrogate by learning systematic CFD-to-experiment discrepancies while preserving its generalization capability and computational efficiency.
Nitin Nagesh Kulkarni, Dheeraj Vemula, Yin Yu +2
Luminary AI 101 S Ellsworth Ave, San Mateo, CA, USA
We introduce GLOBE, a nonlinear neural surrogate for boundary-driven, homogeneous, weakly nonlinear PDEs that draws inductive bias from boundary-element methods and equivariant ML, representing solutions as superpositions of learnable Green's-function-like kernels evaluated from boundary faces to targets. The architecture is translation-, rotation-, and parity-equivariant; discretization-invariant; and units-invariant via rigorous nondimensionalization. On AirFRANS (steady incompressible RANS over NACA airfoils), a compact ~117k-parameter GLOBE model has lower mean-squared error than published values for the strongest reported ML surrogates on both the interpolation and low-data splits, in single-run comparisons under the released AirFRANS normalization. The model evaluates fields at arbitrary points, handles non-watertight meshes, and generalizes under Reynolds number and angle-of-attack distribution shifts. We further introduce a hierarchical Barnes-Hut acceleration that exploits the kernel's far-field decay to reduce evaluation cost from quadratic to near-linear in the number of boundary elements, with a tunable opening angle that recovers exact dense evaluation in the zero-angle limit. Quantitative validation focuses on 2D AirFRANS; on 3D DrivAerML cars, GLOBE predicts surface and volume fields on the full meshes, and we compare it with a matched Transolver baseline. Together, these results show that boundary-element-inspired inductive biases can substantially improve the accuracy of ML surrogates for boundary-driven PDEs while scaling to large 3D meshes.