Neural surrogates for external vehicle aerodynamics are fast alternatives to physics-based computational fluid dynamics (CFD) simulations. Their foundational limitation is a lack of zero-shot generalization across vehicle classes, which results in an expensive, custom training campaign for each new vehicle design. Similar to pretrained models in other fields, a pretrained surrogate should adapt to a new vehicle class using only a few simulations and stay accurate on its pretraining domain. No prior study has systematically tested this transfer behavior. We conduct a large-scale, controlled benchmarking study on five DrivAerML-pretrained surrogates across four independent target datasets. These datasets differ from the source in geometry, CFD solver, turbulence model, and flow conditions. We find that fine-tuning these surrogates outperforms training from scratch in fewer than half of the comparisons and lowers the accuracy of the surrogate on its source dataset. We introduce GeoReplay, a geometry- and gradient-aligned source replay that keeps source simulations resembling the target in training and applies their updates only when they agree with the target gradient. A risk bound separates geometric coverage from differences in the simulated physics, and we prove that every sufficiently small GeoReplay step decreases the target training objective. GeoReplay lowers the source error of fine-tuning in 77 of 80 settings and matches its target accuracy.
Figures & tables
Figure 1: GeoReplay (1) weights source simulations by Chamfer distance to target training geometries, (2) fine-tunes on target and source batches, and (3) adds a source gradient only if it aligns with the target gradient.
Table 1: The source dataset and the four target datasets. Each render shows one simulation from the dataset, coloured by its time-averaged surface pressure on a scale set separately for each dataset. Beyond geometry, the datasets differ in solver, turbulence model, scale, freestream speed, and Reynolds number. Each Reynolds number uses the length scale in parentheses. – marks unreleased values, and ∗ are derived from the released flow conditions. Values are from the dataset papers and cards cited in Section 5.1 .
Figure 2: Source-to-target pairwise geometric distances between every DrivAerML training geometry and every usable geometry in each target dataset. The most distant DrivAerML–target pair for each target dataset, ordered by increasing distance.
Source error eS
Target error eT
Geometric mean eSeT
T
Surrogate
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
DrivAer Net++
TSol
0.042
0.736
0.350
0.096
0.290
1.020
0.215
0.197
0.197
0.197
0.207
0.398
0.263
0.137
0.239
GeoT
0.037
0.764
0.199
0.060
0.063
1.271
0.202
0.205
0.206
0.208
0.217
0.393
0.202
0.111
0.114
DoM
0.186
0.830
0.461
0.135
0.232
0.794
0.242
0.214
0.215
0.214
0.384
0.448
0.314
0.170
0.223
XMGN
0.278
0.719
0.499
0.392
0.371
1.194
0.234
0.226
0.221
0.221
0.576
0.410
0.336
0.294
0.286
GINO
0.268
0.777
0.555
0.360
0.432
0.865
0.283
0.283
0.286
0.284
0.481
0.469
0.397
0.321
0.350
Table 2: Pressure errors of surrogates pretrained on DrivAerML ( S ) and adapted with β=10 training simulations of each target dataset T . The left block gives the source error eS of MTS , the middle block the target error eT of MTT , and the shaded block their geometric mean eSeT . ZS is the pretrained model without adaptation, so its eS does not depend on T . SC trains from scratch on T , FT fine-tunes, GR is GeoReplay , and sGR is sparse GeoReplay . In each block, bold marks the lowest value in each row. See App. B and C for the standard deviation over three seeds and for other quantities and budgets.
DrivAerNet++
SHIFT-SUV
AhmedML
WindsorML
Variant
Weights
Update
eS↓
eT↓
eSeT↓
eS↓
eT↓
eSeT↓
eS↓
eT↓
eSeT↓
eS↓
eT↓
eSeT↓
Transolver
FT (no replay)
–
–
0.350
0.197
0.263
0.414
0.097
0.200
0.272
0.165
0.212
0.333
0.255
0.292
Plain replay
Uniform
Plain
0.072
0.200
0.120
0.070
0.113
0.089
0.071
0.182
0.114
0.071
0.254
0.134
Weighting only
Geometry
Plain
0.071
0.201
0.120
0.069
0.113
0.088
0.074
0.186
0.117
0.074
0.263
0.139
Alignment only
Uniform
Aligned
0.098
0.197
0.139
0.079
0.097
0.087
0.076
0.162
0.111
0.076
0.261
0.141
Table 3: Pressure errors for GeoReplay component ablation at β=10 . The replay variants cross uniform or geometry-weighted source sampling with a plain update gT+λgS or the aligned update of Equation 3 . A-GEM ( Chaudhry et al., 2019 ) , L2-SP ( Li et al., 2018 ) , and EWC ( Kirkpatrick et al., 2017 ) are continual-learning baselines; L2-SP and EWC use the regularization strength in {0.01,0.1,1} with the lowest target validation error. Errors eS , eT , and their geometric mean (shaded) are as in Table 2 . Within each surrogate, bold marks the lowest value in each column and underline the second lowest.
Figure 3: Geometric mean eSeT as a function of the target training budget β , for pressure (top) and wall shear stress (bottom), over the same surrogates in every panel. ZS is the pretrained model without adaptation.
T
TSol
GeoT
DoM
XMGN
GINO
DrivAerNet++
FT
0.1314
0.0406
0.1087
0.1697
0.2065
GR
0.0398
0.0189
0.0287
0.1737
0.1978
SHIFT-SUV
FT
0.0652
0.0577
0.1301
0.1293
0.1577
GR
0.0156
0.0134
0.0302
0.0970
0.1557
AhmedML
FT
0.0662
0.0618
0.1973
0.3470
0.7981
GR
0.0435
0.0313
0.0686
0.3460
0.1655
Table 4: Drag geometric mean eSeT at β=10 for fine-tuning (FT) and GeoReplay (GR), with errors as in Table 2 . Bold marks the lower of each pair. Appendix Table 9 gives both errors and every budget.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8
T
Error
Method
TSol ↓
GeoT ↓
DoM ↓
XMGN ↓
GINO ↓
DrivAerNet++
eS
ZS
0.042 ±0.002
0.037 ±0.002
0.186 ±0.009
0.278 ±0.009
0.268 ±0.003
SC
0.736 ±0.005
0.764 ±0.002
0.830 ±0.004
0.719 ±0.003
0.777 ±0.017
FT
0.350 ±0.002
0.199 ±0.005
0.461 ±0.006
0.499 ±0.010
0.555 ±0.005
GR
0.096 ±0.004
0.060 ±0.000
0.135 ±0.004
0.392 ±0.009
0.360 ±0.007
sGR
0.290 ±0.003
0.063 ±0.001
0.232 ±0.006
0.371 ±0.008
0.432 ±0.007
eT
ZS
1.020 ±0.030
1.271 ±0.037
0.794 ±0.042
1.194 ±0.055
0.865 ±0.040
Appendix
Table 6: Table 2 transposed, with the standard deviation over three training seeds. Columns are surrogates, and rows give the source error eS , the target error eT , and their geometric mean (shaded) for each method and target dataset T . Values are those of Table 2 (mean) with subscripts of the standard deviation over all seeds. Within each surrogate, target dataset, and error, bold marks the lowest of the five methods.
Source error eS
Target error eT
Geometric mean eSeT
T
β
Surrogate
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
DrivAerNet++
1
TSol
0.042
0.789
0.313
0.231
0.306
1.020
0.504
0.347
0.347
0.348
0.207
0.630
0.329
0.283
0.326
GeoT
0.037
0.752
0.166
0.092
0.119
1.271
0.381
0.375
0.376
0.375
0.217
0.535
0.249
0.186
0.211
DoM
0.186
0.804
0.422
0.348
0.396
0.794
0.403
0.329
0.330
0.329
0.384
0.569
0.373
0.339
0.361
XMGN
0.278
0.726
0.479
0.463
0.449
1.194
0.393
0.373
0.368
0.367
0.576
0.534
0.423
0.412
0.406
GINO
0.268
0.764
0.511
0.481
0.498
0.865
0.509
0.404
0.404
0.404
0.481
0.624
0.454
0.441
0.449
Appendix
Table 7: Pressure errors of GeoReplay across target datasets T , target training budgets β and surrogates, laid out as Table 2 . The blocks give the source error eS of MTS on the 48 designs of the official DrivAerML validation split, the target error eT of MTT on the target test simulations, and their geometric mean (shaded). ZS denotes the pretrained model without adaptation, so its errors do not depend on β . In each block, bold marks the lowest value in each row.
Source error eS
Target error eT
Geometric mean eSeT
T
β
Surrogate
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
DrivAerNet++
1
TSol
0.079
0.899
0.451
0.313
0.449
1.155
0.709
0.467
0.467
0.468
0.302
0.798
0.459
0.382
0.458
GeoT
0.070
0.948
0.283
0.152
0.194
1.019
0.477
0.492
0.492
0.493
0.267
0.673
0.373
0.274
0.310
DoM
0.280
1.248
0.550
0.472
0.525
1.091
0.512
0.445
0.446
0.445
0.553
0.799
0.495
0.459
0.483
XMGN
0.400
1.116
0.647
0.632
0.601
1.193
0.473
0.458
0.455
0.455
0.691
0.727
0.544
0.536
0.523
GINO
0.482
0.978
0.943
0.890
0.918
1.208
0.627
0.541
0.541
0.541
0.763
0.783
0.714
0.694
0.705
Appendix
Table 8: Wall shear stress errors of GeoReplay across target datasets T , target training budgets β and surrogates, laid out as Table 2 . The blocks give the source error eS of MTS on the 48 designs of the official DrivAerML validation split, the target error eT of MTT on the target test simulations, and their geometric mean (shaded). ZS denotes the pretrained model without adaptation, so its errors do not depend on β . In each block, bold marks the lowest value in each row.
Source error eS
Target error eT
Geometric mean eSeT
T
β
Surrogate
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
DrivAerNet++
1
TSol
0.0107
0.5594
0.3840
0.4019
0.3902
0.2446
0.1088
0.1297
0.1303
0.1328
0.0512
0.2467
0.2232
0.2288
0.2276
GeoT
0.0084
0.2882
0.0323
0.0647
0.0884
0.5001
0.1134
0.2248
0.2242
0.2214
0.0648
0.1808
0.0852
0.1205
0.1399
DoM
0.0164
1.0640
0.3589
0.3593
0.3615
0.8529
0.1654
0.0732
0.0725
0.0728
0.1181
0.4195
0.1621
0.1614
0.1622
XMGN
0.0893
0.9603
0.2421
0.2429
0.2615
0.3564
0.1939
0.1055
0.1049
0.1044
0.1784
0.4316
0.1598
0.1596
0.1653
GINO
0.1177
0.6088
0.2805
0.2845
0.2822
0.2119
0.2178
0.1596
0.1594
0.1595
0.1579
0.3641
0.2116
0.2129
0.2121
Appendix
Table 9: Drag errors of GeoReplay across target datasets T , target training budgets β and surrogates, laid out as Table 2 . The blocks give the source error eS of MTS on the 48 designs of the official DrivAerML validation split, the target error eT of MTT on the target test simulations, and their geometric mean (shaded). ZS denotes the pretrained model without adaptation, so its errors do not depend on β . In each block, bold marks the lowest value in each row.
Source error eS
Target error eT
Geometric mean eSeT
T
β
Surrogate
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
ZS ↓
SC ↓
FT ↓
GR ↓
sGR ↓
DrivAerNet++
1
TSol
0.256
9.188
2.352
1.253
2.357
3.285
1.701
0.500
0.466
0.483
0.917
3.954
1.085
0.764
1.067
GeoT
0.275
2.527
1.146
0.326
0.280
3.902
2.370
5.693
5.281
5.935
1.035
2.447
2.554
1.313
1.289
DoM
1.604
8.788
3.459
3.776
3.594
14.903
1.660
4.365
4.155
4.336
4.889
3.820
3.886
3.961
3.947
XMGN
2.300
17.559
11.439
11.381
11.298
57.404
3.843
2.886
3.028
2.945
11.490
8.215
5.746
5.870
5.768
GINO
0.988
5.697
0.913
0.922
0.911
4.065
2.269
2.680
2.684
2.690
2.004
3.595
1.564
1.573
1.565
Appendix
Table 10: Lift errors of GeoReplay across target datasets T , target training budgets β and surrogates, laid out as Table 2 . The blocks give the source error eS of MTS on the 48 designs of the official DrivAerML validation split, the target error eT of MTT on the target test simulations, and their geometric mean (shaded). ZS denotes the pretrained model without adaptation, so its errors do not depend on β . In each block, bold marks the lowest value in each row.
Pressure relative L2
Wall shear stress relative L2
Target dataset
Surrogate
Mean ↓
ZS ↓
FT ↓
SC ↓
SAV (%) ↑
Mean ↓
ZS ↓
FT ↓
SC ↓
SAV (%) ↑
DrivAerNet++
TSol
0.771
1.020
0.197
0.215
+8.3
0.893
1.155
0.300
0.315
+4.7
GeoT
1.271
0.205
0.202
−1.7
1.019
0.307
0.315
+2.6
DoM
0.794
0.214
0.242
+11.6
1.091
0.317
0.342
+7.2
XMGN
1.194
0.226
0.234
+3.4
1.193
0.338
0.335
−0.8
GINO
0.865
0.283
0.283
−0.1
1.208
0.386
0.380
−1.6
Appendix
Table 11: At β=10 target training simulations, mean test error for the mean-field baseline (Mean), zero-shot source-only prediction (ZS), fine-tuning (FT), and training from scratch (SC), with SAV=1−FT/SC , on the target test simulations. Blue and orange shade positive and negative SAV estimates, and gray marks ∣SAV∣<1% . Bold / underlined mark the lowest/second-lowest errors per SC/FT:target results.
β
FT
SC
SAV
1
0.653
0.882
+26.0
4
0.299
0.325
+7.8
10
0.173
0.126
−37.4
20
0.108
0.054
−98.4
40
0.084
0.058
−45.1
150
0.067
0.066
−1.1
Appendix
Table 12: AB-UPT on AhmedML: mean pressure L2 test error of fine-tuning (FT) and training from scratch (SC) for each target training budget β , and SAV=1−FT/SC . There is no β=80 run.
Figure 5: Each row shows the test geometry nearest the median per-simulation pressure advantage of fine-tuning over training from scratch, at β=80 ( β=79 for SHIFT-SUV). Bold marks the lowest relative L2 error among SC, FT, and GR, and dashed ellipses mark the same surface region on their maps.
Dataset
Model
Ref
Mix
Δ
DrivAerML
Transolver
0.055
0.048
− 12%
GeoTransolver
0.046
0.040
− 13%
DrivAerNet++
Transolver
0.175
0.172
− 1%
GeoTransolver
0.169
0.166
− 2%
SHIFT-SUV
Transolver
0.050
0.051
+1%
GeoTransolver
0.047
0.048
+2%
Appendix
Table 13: In-domain evaluation at equal exposure. Pressure L2 error on the test designs of each dataset (the official split where a dataset publishes one) for the reference trained on that dataset only (Ref) and the model trained on all five datasets (Mix), with Δ=Mix/Ref−1 ; negative Δ means that the mixture is more accurate. The mixture draws each dataset as often as its reference. All models train in bfloat16 precision and are evaluated after the last of 501 epochs. Transolver values are the mean of two seeds (spread at most 4%).
Deploying Scientific Machine Learning surrogates in industrial CFD workflows requires adapting pretrained models to new vehicle families without large datasets; yet whether geometric representations learned by a geometry encoder transfer to topologically distinct shapes remains unvalidated. We address this through leave-one-family-out experiments on a 61.47M-parameter Transformer surrogate (AB-UPT) pretrained on four vehicle families (411 external aerodynamics cases) and adapted to the held-out fifth with only 20 samples. Three strategies are compared: Full Fine-Tuning (FFT), Lightweight Fine-Tuning (LFT), and Low-Rank Adaptation (LoRA). The central finding is that pretrained geometry encoders learn transferable representations, but the adaptation mechanism determines whether they can be exploited. FFT destabilizes as 61.47M unconstrained parameters overfit to 20 samples (R^2=0.40); LFT fails because the frozen encoder cannot represent unseen shapes (R^2<0). LoRA resolves both: rank-constrained adapters injected into all layers regularize the loss landscape while preserving pretrained features, achieving R^2=0.85+/-0.02 across all five families with 50% lower force RMSE than FFT and 28% lower pointwise field errors. LoRA also outperforms from-scratch training using 3x more target-family data, eliminating the need for large per-family datasets. These results recast LoRA from a memory-saving convenience into a convergence enabler for geometry transfer: a shared backbone paired with lightweight per-family adapters trainable in hours from minimal data.
Neural surrogates for computational fluid dynamics (CFD) have the potential to greatly enhance engineering innovation through accelerating simulation. However, the primary limitation for neural surrogates is the lack of generalization to geometries and applications beyond the training set, which is significant given the diversity of engineering scenarios. Currently, this is addressed by generating a new dataset for a specific application; however, this requires running costly numerical solvers. In this work, we take a step toward addressing this by studying neural surrogates trained across different geometries, boundary conditions, and fidelities. We find that cross-domain pretraining improves zero- and few-shot performance on held-out datasets relative to both training from scratch and transferring from domain-specific experts. In particular, finetuning a pretrained, cross-domain model can achieve 2-3x lower errors at the same sample size and use 8x fewer samples to achieve the same error, compared to training from scratch. This benefit is architecture agnostic and improves with model size and pretraining dataset diversity. Furthermore, we study how and why cross-domain pretraining works in CFD surrogates, and find that simply pooling steady-state datasets is both sufficient and effective. Given the high cost of generating CFD data, leveraging existing datasets through cross-domain pretraining will likely be a valuable strategy as future surrogates expand to tackle new problems and use cases.
Anthony Zhou, Amir Barati Farimani, Shirley Ho +1
Carnegie Mellon University · New York University · Polymathic AI +4
Computational Fluid Dynamics (CFD) is central to race-car aerodynamic development, yet its cost -- tens of thousands of core-hours per high-fidelity evaluation -- severely limits the design space exploration feasible within realistic budgets. AI-based surrogate models promise to alleviate this bottleneck, but progress has been constrained by the limited complexity of public datasets, which are dominated by smoothed passenger-car shapes that fail to exercise surrogates on the thin, complex, highly loaded components governing motorsport performance. This work presents three primary contributions. First, we introduce a high-fidelity RANS dataset built on a parametric LMP2-class CAD model and spanning six operating conditions (map points) covering straight-line and cornering regimes, generated and validated by aerodynamics experts at Dallara to preserve features relevant to industrial motorsport. Second, we present the Gauge-Invariant Spectral Transformer (GIST), a graph-based neural operator whose spectral embeddings encode mesh connectivity to enhance predictions on tightly packed, complex geometries. GIST guarantees discretization invariance and scales linearly with mesh size, achieving state-of-the-art accuracy on both public benchmarks and the proposed race-car dataset. Third, we demonstrate that GIST achieves a level of predictive accuracy suitable for early-stage aerodynamic design, providing a first validation of the concept of interactive design-space exploration -- where engineers query a surrogate in place of the CFD solver -- within industrial motorsport workflows.
Nicholas Thumiger, Andrea Bartezzaghi, Mattia Rigotti +5