betaFeedback

WorksComputer Science

Measuring and Improving Transfer of Neural Surrogates for Vehicle Aerodynamics

Abstract

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.

Introduction

Aerodynamic drag and lift coefficients govern vehicle range, stability, and fuel economy. Design engineers evaluate hundreds of geometries against these coefficients before converging on a design. Drag and lift are surface integrals of the pressure and wall shear stress. Computing these quantities requires numerically solving the Navier–Stokes equations around the vehicle, which is a task traditionally done by high-fidelity CFD solvers. A single simulation can take hours to converge. Neural surrogates have emerged as computationally efficient alternatives. Given a vehicle geometry, these surrogates can predict surface and volume flow fields in milliseconds to seconds [40, 1, 33, 12, 29], which makes wide design sweeps tractable (Anantham et al., 2024; Ashton et al., 2024c,a; Elrefaie et al., 2024). This speed comes at an upfront cost. Each surrogate has to be trained on many CFD runs of the vehicle class, and every new class, e.g., a different body style, segment, or program, requires repeating this expensive training process.

A pretrained aerodynamic surrogate should be able to learn the underlying physics and adapt to a new vehicle class using only a few simulations while staying accurate on its source dataset, similar to how models pretrained at scale in other domains generalize to unseen data and adapt to new tasks with little data [9]. This generalization is critical in practice since each new target domain typically comes with only a handful of CFD runs. Generating more simulations is precisely the cost that pretraining is meant to avoid [42].

Most automotive surrogates are trained and evaluated with the same vehicle dataset [7, 17, 38, 18], whereas practitioners already apply pretrained surrogates across a broad range of geometries and application areas [13]. Although there are only a handful of publicly available CFD datasets, together they provide a diverse testbed for studying transfer across vehicle classes. These datasets span variations that a deployed surrogate would see in practice, e.g., parametrically varied notchback road cars (DrivAerML) [7]; fastback, notchback, and estateback production-style vehicles (DrivAerNet++) [17]; quarter-scale sport utility vehicles (SHIFT-SUV) [26]; simplified Ahmed reference bodies (AhmedML) [6]; and square-back reference bodies (WindsorML) [5]. Table 1 summarizes how these datasets additionally differ in CFD solver, turbulence model, scale, and flow conditions.

DrivAerMLDrivAerNet++SHIFT-SUVAhmedMLWindsorML
Source/TargetSourceTarget, road carTarget, road carTarget, bluff bodyTarget, bluff body
GeometryDrivAer notchbackFast-, notch-, estatebackAeroSUVAhmed bodyWindsor square-back
Shape parameters16261067
SolverOpenFOAM v2212OpenFOAM v11Luminary CloudOpenFOAM v2212Volcano ScaLES
Turbulence modelSA-σ\sigma-DDESSteady RANS, kk-ω\omega SSTDDESSA-DDESWMLES
Surface fieldsTime-averagedSteadyTime-averagedTime-averagedTime-averaged
Mesh cells≈\approx160M≈\approx24M≈\approx45M15–20M>280>280M
ScaleFullFullQuarterLaboratoryLaboratory
Freestream speed38.9 m/s30 m/s50 m/s1 m/s40 m/s
Reynolds number7.2×1067.2 \times10^{6} (wheelbase)8.48.4–10.1×10610.1 \times10^{6} (length)3.9×1063.9 \times10^{6} (length)∗^{*}2.78×1062.78 \times10^{6} (length)2.9×1062.9 \times10^{6} (length)
Released simulations5008,00099 (sample)500350

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.

In this work, we adopt them as our transfer learning testbed and aim to answer the following question: How can a pretrained aerodynamic surrogate adapt to a new vehicle class from a few simulations while keeping its accuracy on the source dataset? Answering this question raises three challenges. (i) Attribution. Improvement after fine-tuning mixes the value of pretraining with the value of the target simulations themselves. A credible comparison must hold the architecture, data, and selection protocol fixed. (ii) Coverage versus physics. A target geometry close to the source can still differ in solver, mesh, and flow conditions, so geometric similarity alone does not determine transfer. (iii) Conflicting source updates. Source information that helps one target can hurt another, so adaptation must detect when a source update opposes the target objective.

Our approach. For (i), we compare each fine-tuned model with the same architecture trained from scratch on the same target simulations, and we evaluate both on the source and on the target (§ 3). For (ii), we bound the target risk by a geometric coverage term plus a physics-shift term, which explains why geometric similarity alone cannot guarantee transfer. For (iii), we introduce GeoReplay, a geometry- and gradient-aligned source replay (Figure 1, Algorithm 1, § 4). An entropy-regularized coupling weights source simulations by their proximity to the target training geometries, and a cosine gate keeps a source gradient only when it agrees with the target gradient. Across surrogates, target datasets, and budgets, GeoReplay keeps the target accuracy of fine-tuning while substantially reducing its error on the source dataset (§ 5).

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.

In summary, our contributions are as follows.

  • We introduce GeoReplay, a geometry- and gradient-aligned source replay method that adapts a pretrained surrogate to a new target while preserving its source accuracy (§ 4).

  • We bound the target risk of geometry-weighted replay by separating geometric coverage from physics shift, and prove that GeoReplay steps descend the target training objective (§ 4).

  • We present the first large-scale study, to our knowledge, of transfer across independent automotive

CFD datasets that span distinct geometries, CFD solvers, turbulence models, and Reynolds numbers, evaluating jointly on their source and target domains (§5).

Related Work

Aerodynamic surrogate benchmarks. Prior studies do not measure transfer between independent CFD campaigns. Public CFD benchmarks ((Ashton et al., 2024c,b,a; Elrefaie et al., 2024; Luminary Cloud, 2025)) mostly assess surrogates within one dataset ([38]) or under category shifts inside it. For example, CarBench ([18]) benchmarks various surrogates on different car geometries in DrivAerNet++. A limitation of this study is that it does not consider changes in the numerical solver or turbulence modeling. Surrogate architectures also continue to evolve, for example toward predicting from raw point-cloud geometry without a simulation mesh ([19]).

Negative transfer and transferability. Negative transfer is typically assessed by comparing an adapted model with a target-only baseline ([39, 20]), and metrics fitted from target labels can estimate transferability before adaptation ([30, 31]). Continual learning limits forgetting of the earlier task by regularizing toward the pretrained parameters, as in EWC ([22]) and L2-SP ([23]), or by replaying earlier data, as in experience replay ([35]) and A-GEM ([11]).

Transfer of scientific surrogates. For scientific surrogates, transfer depends on how closely source and target physics match, and neural operators ([24]) degrade out of distribution ([37, 27, 28]); parameter-efficient methods adapt PDE surrogates on lower-dimensional benchmarks instead of three-dimensional vehicle aerodynamics ([8, 10]). Cross-domain pretraining of steady-state CFD surrogates over geometries, boundary conditions and simulation fidelities lowers the error of fine-tuned models relative to training from scratch ([43]); Appendix H compares surrogates trained on a mixture of the five vehicle datasets with surrogates trained on each dataset alone.

Problem Definition

Let DSD_S and DTD_T denote the source and target datasets, respectively. For an architecture fθf_\theta, we compare a model initialized from parameters θsrc\theta_{\mathrm{src}} learned on DSD_S with the same architecture initialized randomly. Both models train on β\beta simulations from DTD_T and use the same validation simulations for checkpoint selection. We evaluate both models on the same held-out simulations. We also evaluate the source-pretrained model directly on the target test simulations (zero-shot prediction). We use zero-shot error to measure initial generalization, its reduction after fine-tuning to measure adaptation, and the fine-tuned–from-scratch comparison to measure the realized value of pretraining at fixed β\beta. For a method M, a subscript denotes the training dataset and a superscript the test dataset. The errors eSe_S and eTe_T denote the errors of MTSM^S_T and MTTM^T_T, respectively. Since a reusable surrogate must stay accurate on both datasets, we propose to use the geometric mean eSeT\sqrt{e_S e_T} of both errors, which is low only when both errors are.

Distance to the closest training geometry. We follow the evaluation structure from recent automotive aerodynamic analyses that use Chamfer distance to quantify geometric diversity and relate prediction error to the closest training geometry ((PhysicsX Research, 2026)). Let P(g)P(g) contain area-weighted surface points from geometry gg after reconciling dataset axes, centering the body, and uniformly normalizing all coordinates by body length. For two geometries, we compute the symmetric Chamfer distance

dC(g,g′)=12[1∣P(g)∣∑x∈P(g)min⁡y∈P(g′)∥x−y∥2+1∣P(g′)∣∑y∈P(g′)min⁡x∈P(g)∥y−x∥2].(1)d_C(g,g') = \frac{1}{2}\left[\frac{1}{|P(g)|}\sum_{x\in P(g)}\min_{y\in P(g')} \lVert x-y\rVert_2 + \frac{1}{|P(g')|}\sum_{y\in P(g')}\min_{x\in P(g)} \lVert y-x\rVert_2\right]. \tag*{(1)}

For source DrivAerML training geometries SS, the distance to the closest training geometry is dS(g)=min⁡g′∈SdC(g,g′)d_S(g)=\min_{g'\in S}d_C(g,g'). It is dimensionless and requires no target CFD field. We use all unordered pairs within each dataset to characterize internal geometric diversity. We separately compare all DrivAerML–target pairs. For each fixed target test geometry, we compare its distance to the closest DrivAerML pretraining geometry with its distance to the closest geometry in the corresponding target training split. We measure this distance before any training. See Appendix A for the computation details.

GeoReplay: Geometry- and Gradient-Aligned Source Replay

We observe two obstacles to reliable transfer. The distance to the source geometry set can characterize a model’s initial geometric extrapolation, but it cannot determine whether pretraining remains useful after fine-tuning. Replaying arbitrary source simulations during adaptation can bias training toward source-specific solvers, meshes, operating conditions, or turbulence models. We use geometry to decide which source simulations are relevant and the target objective to decide whether their gradients are useful.

Geometry-Weighted Source Distribution

We denote a simulation as z=(g,y)z=(g,y), where gg denotes a surface geometry and yy contains its CFD fields. Let P^TG=β−1∑i=1βδgiT\widehat{P}_T^G=\beta^{-1}\sum_{i=1}^{\beta}\delta_{g_i^T} be the empirical target-geometry distribution, let Cij=dC(giT,gjS)C_{ij}=d_C(g_i^T,g_j^S), and let uj=1/nSu_j=1/n_S. We compute an entropy-regularized target-to-source coupling

Π⋆=arg⁡min⁡Π≥0Π1=β−11⟨C,Π⟩+τDKL(Π∥β−11u⊤),(2)\Pi^{\star}=\mathop{\arg\min}_{\Pi\geq0\atop\Pi\mathbf{1}=\beta^{-1}\mathbf{1}}\langle C,\Pi\rangle+\tau D_{\mathrm{KL}}\left(\Pi\middle\|\beta^{-1}\mathbf{1}u^{\top}\right), \tag*{(2)}
w⋆=Π⋆⊤1,w^{\star}=\Pi^{\star\top}\mathbf{1},

where dCd_C denotes the body-length-normalized Chamfer distance from Equation (1) and τ>0\tau>0 controls concentration. This entropy-regularized transport objective [15] has the row-wise solution

Πij⋆=β−1exp⁡(−Cij/τ)∑kexp⁡(−Cik/τ).\Pi_{ij}^{\star}=\beta^{-1}\frac{\exp(-C_{ij}/\tau)}{\sum_k\exp(-C_{ik}/\tau)}.

Its source marginal w⋆w^{\star} gives every target geometry equal mass while favoring nearby source geometries. Equation (2) uses the target training geometries but does not inspect their CFD fields.

Target-Aligned Replay Updates

Let FT(θ)F_T(\theta) denote the target empirical risk and let FS,w⋆(θ)F_{S,w^{\star}}(\theta) denote the source risk under w⋆w^{\star}. At one optimization step, we compute a target gradient gTg_T and a geometry-weighted source gradient gSg_S. We retain only the component of the source update that agrees with the target descent direction:

a(gT,gS)=max⁡(0,⟨gT,gS⟩∥gT∥∥gS∥+ϵ),a(g_T,g_S)=\max\left(0,\frac{\langle g_T,g_S\rangle}{\lVert g_T\rVert\lVert g_S\rVert+\epsilon}\right),
gR=a(gT,gS)∥gT∥∥gS∥+ϵgS.(3)g_R=a(g_T,g_S)\frac{\lVert g_T\rVert}{\lVert g_S\rVert+\epsilon}g_S. \tag*{(3)}

The update is θ+=θ−η(gT+λgR)\theta^{+}=\theta-\eta(g_T+\lambda g_R), where λ≥0\lambda\geq0 sets the replay strength. The cosine gate in Equation (3) is the gradient-similarity rule of Du et al. [16] for auxiliary losses, to which we add a rescaling to the target gradient norm. Unlike gradient surgery [41], it leaves the target gradient unchanged. GEM and A-GEM [25, 11] constrain each update with the gradient of an episodic memory of earlier-task samples; GeoReplay uses the source simulations as this memory, samples them by geometry, and drops conflicting source gradients instead of projecting. If no sampled source batch has positive alignment, gR=0g_R=0 and the method reduces to ordinary target fine-tuning. We summarize the full GeoReplay procedure in Algorithm 1.

Algorithm 1 GeoReplay

Require: θsrc\theta_{\mathrm{src}}, DSD_S, DTD_T, replay strength λ\lambda, entropy weight τ\tau

  1. Compute pairwise geometry costs dC(giT,gjS)d_C(g_i^T, g_j^S)

  2. Solve Equation (2) for source weights w∗w^*

  3. θ←θsrc\theta\leftarrow\theta_{\mathrm{src}}

  4. while the target validation criterion improves do

  5. Sample a target batch MT\mathcal{M}_T and a source batch MS∼w∗\mathcal{M}_S \sim w^*

  6. gT←∇θF^MT(θ)g_T \leftarrow\nabla_\theta\widehat{F}_{\mathcal{M}_T}(\theta); gS←∇θF^MS(θ)g_S \leftarrow\nabla_\theta\widehat{F}_{\mathcal{M}_S}(\theta)

  7. Compute gRg_R using Equation (3)

  8. θ←θ−η(gT+λgR)\theta\leftarrow\theta- \eta(g_T + \lambda g_R)

  9. end while

  10. return the checkpoint with minimum target validation error

What the Two Alignment Steps Guarantee

The geometry weights control one identifiable part of source–target discrepancy. They cannot remove changes in the conditional CFD solution induced by a different solver, mesh, operating condition, or turbulence model. Proposition 1 makes this separation explicit. For a fixed θ\theta, define the conditional risks mD(g)=ED[ℓ(fθ(g),Y)∣G=g]m_D(g) = \mathbb{E}_D[\ell(f_\theta(g), Y) \mid G = g] for D∈{S,T}D \in\{S,T\}.

Proposition 1 (Geometry-conditioned risk discrepancy). Assume that ∣mT(g)−mT(g′)∣≤LGdC(g,g′)\lvert m_T(g) - m_T(g') \rvert\le L_G d_C(g,g') and that ∣mT(g)−mS(g)∣≤δphys\lvert m_T(g) - m_S(g) \rvert\le\delta_{\mathrm{phys}} on the support of QwGQ^G_w. Let TdC(P,Q)=inf⁡π∈Π(P,Q)E(G,G′)∼π[dC(G,G′)]\mathcal{T}_{d_C}(P,Q) = \inf_{\pi\in\Pi(P,Q)} \mathbb{E}_{(G,G') \sim\pi}[d_C(G,G')]. Then

∣RT(θ)−RS,w(θ)∣≤LGTdC(PTG,QwG)+δphys.(4)\lvert R_T(\theta) - R_{S,w}(\theta) \rvert\le L_G \mathcal{T}_{d_C}(P_T^G,Q_w^G) + \delta_{\mathrm{phys}}. \tag*{(4)}

The first term decreases when the weighted source geometries cover the target geometries. The residual δphys\delta_{\mathrm{phys}} includes conditional differences that geometry alone cannot identify. This term explains why geometric proximity can coexist with negative transfer. Proposition 1 is an instance of the Wasserstein-type domain adaptation bounds of Redko et al. [34], Courty et al. [14], and Shen et al. [36], with the Chamfer distance as ground metric and the physics shift as the residual term.

Proposition 2 (Target-objective descent). Assume that FTF_T is LL-smooth, gT=∇FT(θ)g_T = \nabla F_T(\theta), and ⟨gT,gR⟩≥0\langle g_T,g_R\rangle\ge0. Let d=gT+λgRd = g_T + \lambda g_R. For any

0<η≤2⟨gT,d⟩L∥d∥2,(5)0 < \eta\le\frac{2\langle g_T,d\rangle}{L\lVert d\rVert^2}, \tag*{(5)}

the update θ+=θ−ηd\theta^+ = \theta- \eta d satisfies FT(θ+)≤FT(θ)F_T(\theta^+) \le F_T(\theta).

The cosine gate enforces the alignment condition by construction. Proposition 2 is a one-step statement about the target training objective; it does not guarantee lower held-out error or improvement over ordinary fine-tuning. See Appendix E for the proofs.

Complexity. Constructing the geometry cost matrix requires O(βnSCd)O(\beta n_S C_d) work, where CdC_d denotes the cost of one geometry comparison, and normalizing the coupling requires O(βnS)O(\beta n_S) work. These computations occur once before adaptation. Each training step adds one source forward and backward pass. The gradient gate requires O(p)O(p) operations for pp trainable parameters. Source and target gradients can be accumulated sequentially, so the method does not need to store both computation graphs at once.

Experiments

Experimental Setup

Datasets. We use five independent CFD datasets, one source dataset and four target datasets (Table 1).

The source contains the 436 DrivAerML road-car simulations used to pretrain the released models [7]. The targets contain 500 DrivAerNet++ production-car simulations [17], 99 SHIFT-SUV quarter-scale sport utility vehicle simulations [26], whose 10 shape parameters and ≈45\approx45M-cell meshes are documented by [3] and [1], 499 complete full-body AhmedML simulations [6], and the 350 published WindsorML square-back simulations [5].

Models. The complete evaluation pairs DrivAerNet++, SHIFT-SUV, AhmedML, and WindsorML with four DrivAerML-pretrained surface surrogates released through PhysicsNeMo-CFD [38]: Transolver (TSol) [40], GeoTransolver (GeoT) [1], DoMINO (DoM) [33], and X-MeshGraphNet (XMGN) [29], and with GINO (Li et al., 2023). GINO has no released DrivAerML checkpoint, so we pretrain it on the same 436-simulation DrivAerML split with its published recipe (Appendix § 13). For reference, the appendix also reports AB-UPT [2] against training from scratch (Table 11). It uses its released DrivAerML checkpoint and requires the cell centers of each geometry’s CFD volume mesh, so we keep it outside the benchmark counts.

Target datasetSurrogatePressure relative L2L_2 Mean ↓ZS ↓FT ↓SC ↓SAV (%) ↑Wall shear stress relative L2L_2 Mean ↓ZS ↓FT ↓SC ↓SAV (%) ↑
DrivAerNet++TSol0.7711.0200.197<u>0.215</u>+8.30.8931.1550.3000.315+4.7
GeoT1.271<u>0.205</u>0.202−1.71.0190.3070.315+2.6
DoM0.7940.2140.242+11.61.0910.3170.342+7.2
XMGN1.1940.2260.234+3.41.1930.338<u>0.335</u>−0.8
GINO0.8650.2830.283−0.11.2080.3860.380−1.6
SHIFT-SUVTSol0.8920.733<u>0.097</u><u>0.103</u>+5.90.9581.422<u>0.179</u>0.172−4.2
GeoT0.8450.0810.067−20.71.2640.1330.100−33.6
DoM0.6870.1190.134+11.11.2640.2060.254+19.0
XMGN1.1100.117<u>0.103</u>−13.71.9370.198<u>0.167</u>−18.5
GINO0.5820.2390.221−8.31.7310.3880.365−6.3
AhmedMLTSol0.8811.041<u>0.165</u>0.136−21.60.7000.876<u>0.133</u><u>0.116</u>−14.3
GeoT0.9880.1480.124−19.10.8000.1150.108−6.2
DoM1.2420.2560.243−5.50.7930.1730.172−0.6
XMGN2.0360.2920.329+11.30.9100.2080.234+11.1
GINO1.0430.4860.590+17.60.7750.3690.443+16.6
AB-UPT†0.9710.173<u>0.126</u>−37.40.8270.1520.124−22.7
WindsorMLTSol0.6940.794<u>0.255</u>0.427+40.30.5470.877<u>0.193</u>0.296+34.8
GeoT0.7440.2440.173−40.90.7930.1750.131−32.9
DoM1.5190.2950.279−5.61.3440.2180.204−6.6
XMGN0.9000.3220.329+2.20.8890.2220.237+6.2
GINO0.7900.4870.470−3.70.8090.3630.352−3.3

Table 11. At β=10\beta= 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= 1 - \mathrm{FT}/\mathrm{SC}, on the target test simulations. Blue and orange shade positive and negative SAV estimates, and gray marks ∣SAV∣<1%|\mathrm{SAV}| < 1\%. Bold/underlined mark the lowest/second-lowest errors per SC/FT:target results.

Training and evaluation. Each target uses ten validation and ten test simulations, except WindsorML, for which we use the official validation and test splits of the release (35 simulations each). We report the main comparison at β=10\beta= 10 target training simulations, where only a few labeled target geometries are available, and repeat it at β∈{1,40,80}\beta\in\{1, 40, 80\} (Appendix § 10). SHIFT-SUV uses β=79\beta= 79 at the largest budget because only 99 surface simulations are available. Each model trains until its validation error stops improving (early stopping with patience 30, at most 500 epochs), and we select the checkpoint with the lowest validation pressure and wall shear stress error. Fine-tuned and from-scratch runs use the same PhysicsNeMo model implementation and split. We train every comparison with seed 0 and repeat the Transolver and GeoTransolver comparisons with seeds 1 and 2 (Appendix § 9). Appendix F contains the training recipes and dataset conversions.

For pressure and wall shear stress, we compute relative L2L_2 error on each complete surface and average it over the target test simulations for eTe_T and over the 48 designs of the official DrivAerML validation split for eSe_S. The wall shear stress norm includes all three vector components. We integrate the predicted pressure and wall shear stress over the CFD surface to obtain drag and report mean relative drag error. Because each error is relative to the magnitude of the CFD field or force, it is dimensionless and comparable across datasets with different units and scales.

Results

GeoReplay preserves source accuracy at matched target accuracy. We report pressure errors at β=10\beta= 10 in Table 2. Fine-tuning lowers the target error of every surrogate but raises its DrivAerML error far above zero-shot prediction, from 0.042 to 0.350 for Transolver adapted to DrivAerNet++. GeoReplay keeps the target error of fine-tuning within 0.01 in 19 of 20 surrogate and target pairs while lowering the DrivAerML error by a median factor of 3.3, to 0.096 in the same example, and it has the lowest geometric mean in 18 of the 20 rows. Across all 80 combinations of surrogate, target dataset, and budget (Appendix § 10), GeoReplay lowers the DrivAerML pressure error of fine-tuning in 77 and of training from scratch in all 80, while its median target error differs from fine-tuning by 0.2%. We report the standard deviation over three training seeds in Appendix Table 6. Across seeds, the standard deviation of each error is at most 9% of its value and that of the geometric mean at most 3%, far smaller than the gap in DrivAerML error between fine-tuning and GeoReplay. Among the adapted methods, GeoReplay has the lowest median geometric mean for every target dataset and budget in Figure 3 except AhmedML pressure at β=1\beta= 1 (sparse GeoReplay, 0.004 lower). As expected, the geometric mean of every adapted method is lower at β=10\beta= 10 than at β=1\beta= 1. Drag shows the same pattern, with a lower drag geometric mean than fine-tuning in 19 of 20 pairs (Table 4). The pattern also holds for wall shear stress and lift, where GeoReplay has a lower geometric mean than fine-tuning in 20 of 20 and 19 of 20 pairs at β=10\beta= 10 (Appendix Tables 8 and 10).

Table 2: Pressure errors of surrogates pretrained on DrivAerML (SS) and adapted with β=10\beta= 10 training simulations of each target dataset TT. The left block gives the source error eSe_S of MTSM^S_T, the middle block the target error eTe_T of MTTM^T_T, and the shaded block their geometric mean eSeT\sqrt{e_S e_T}. ZS is the pretrained model without adaptation, so its eSe_S does not depend on TT. SC trains from scratch on TT, FT fine-tunes, GR is GeoReplay, and sGR is sparse GeoReplay. In each block, bold marks the lowest value in each row. See App. § 9 and App. § 10 for the standard deviation over three seeds and for other quantities and budgets.

TTSurrogateSource error eSe_S: ZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowTarget error eTe_T: ZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowGeometric mean eSeT\sqrt{e_S e_T}: ZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrow
DrivAerNet++TSol0.0420.7360.3500.0960.2901.0200.2150.1970.1970.1970.2070.3980.2630.1370.239
GeoT0.0370.7640.1990.0600.0631.2710.2020.2050.2060.2080.2170.3930.2020.1110.114
DoM0.1860.8300.4610.1350.2320.7940.2420.2140.2150.2140.3840.4480.3140.1700.223
XMGN0.2780.7190.4990.3920.3711.1940.2340.2260.2210.2210.5760.4100.3360.2940.286
GINO0.2680.7770.5550.3600.4320.8650.2830.2830.2860.2840.4810.4690.3970.3210.350
SHIFT-SUVTSol0.0420.7220.4140.0800.2900.7330.1030.0970.0980.0980.1750.2720.2000.0890.168
GeoT0.0370.7390.2360.0570.0570.8450.0670.0810.0820.0820.1770.2230.1380.0680.069
DoM0.1861.2610.4850.1340.1670.6870.1340.1190.1240.1200.3580.4110.2400.1290.141
XMGN0.2780.8140.7140.4450.3961.1100.1030.1170.1110.1110.5560.2890.2890.2220.210
GINO0.2680.7200.9720.4280.5090.5820.2210.2390.2390.2390.3950.3990.4820.3200.349
AhmedMLTSol0.0420.8240.2720.0750.2111.0410.1360.1650.1700.1650.2090.3350.2120.1130.187
GeoT0.0370.8090.2260.0580.0660.9880.1240.1480.1490.1450.1920.3170.1820.0930.098
DoM0.1861.3060.2860.1460.1841.2420.2430.2560.2530.2540.4810.5630.2710.1920.216
XMGN0.2780.9410.3960.3320.3522.0360.3290.2920.2880.2970.7530.5570.3400.3090.323
GINO0.2681.7070.5790.2700.2771.0430.5900.4860.4740.4700.5281.0030.5300.3580.361
WindsorMLTSol0.0421.2270.3330.0760.3100.7940.4270.2550.2530.2560.1820.7240.2920.1390.282
GeoT0.0370.8050.2230.0580.0610.7440.1730.2440.2370.2370.1660.3730.2330.1170.120
DoM0.1861.8080.2960.1440.1691.5190.2790.2950.2910.2940.5320.7100.2950.2050.223
XMGN0.2780.8700.4510.3480.3770.9000.3290.3220.3240.3230.5000.5350.3810.3350.349
GINO0.2680.9010.8670.2710.2800.7900.4700.4870.4840.4840.4600.6510.6500.3620.368

Table 2.

Figure 3. Geometric mean eSeT\sqrt{e_S e_T} as a function of the target training budget β\beta, for pressure (top) and wall shear stress (bottom), over the same surrogates in every panel. ZS is the pretrained model without adaptation.

TTTSolGeoTDoMXMGNGINO
DrivAerNet++ FT0.13140.04060.10870.16970.2065
DrivAerNet++ GR0.03980.01890.02870.17370.1978
SHIFT-SUV FT0.06520.05770.13010.12930.1577
SHIFT-SUV GR0.01560.01340.03020.09700.1557
AhmedML FT0.06620.06180.19730.34700.7981
AhmedML GR0.04350.03130.06860.34600.1655
WindsorML FT0.07090.07130.24350.41750.7215
WindsorML GR0.04370.04000.05370.36680.1351

Table 4. Drag geometric mean eSeT\sqrt{e_S e_T} at β=10\beta= 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.

Component ablation. In Table 3, we separate the source weighting from the update rule for Transolver and GeoTransolver. Plain replay with uniform weights corresponds to experience replay ([35]) on the source simulations. Every replay variant lowers the DrivAerML error of fine-tuning by a factor of 3.3 to 5.8, but only GeoReplay keeps the target error within 0.005 of fine-tuning on every target dataset. The plain update raises the target error by up to 0.042 and the aligned update with uniform weights by up to 0.022, both on SHIFT-SUV, so both components are needed to preserve target accuracy.

Table 3: Pressure errors for GeoReplay component ablation at β=10\beta= 10. The replay variants cross uniform or geometry-weighted source sampling with a plain update gT+λgSg_T + \lambda g_S or the aligned update of Equation (3). A-GEM [11], L2-SP [23], and EWC [22] are continual-learning baselines; L2-SP and EWC use the regularization strength in {0.1,1,1}\{0.1, 1, 1\} with the lowest target validation error. Errors eSe_S, eTe_T, and their geometric mean (shaded) are as in Table 2. Within each surrogate, bold marks the lowest value in each column and <u>underline</u> the second lowest.

VariantWeightsUpdateDrivAerNet++ eS↓e_S \downarroweT↓e_T \downarroweSeT↓\sqrt{e_S e_T}\downarrowSHIFT-SUV eS↓e_S \downarroweT↓e_T \downarroweSeT↓\sqrt{e_S e_T}\downarrowAhmedML eS↓e_S \downarroweT↓e_T \downarroweSeT↓\sqrt{e_S e_T}\downarrowWindsorML eS↓e_S \downarroweT↓e_T \downarroweSeT↓\sqrt{e_S e_T}\downarrow
Transolver
FT (no replay)––0.3500.1970.2630.4140.0970.2000.2720.1650.2120.3330.2550.292
Plain replayUniformPlain<u>0.072</u>0.2000.120<u>0.070</u>0.1130.0890.0710.1820.1140.071<u>0.254</u>0.134
Weighting onlyGeometryPlain0.0710.201<u>0.120</u>0.0690.113<u>0.088</u>0.0740.1860.1170.0740.2630.139
Alignment onlyUniformAligned0.0980.1970.1390.0790.0970.0870.0760.1620.1110.0760.2610.141
A-GEMUniformProjection0.132<u>0.198</u>0.1610.086<u>0.097</u>0.0910.077<u>0.159</u><u>0.111</u>0.0790.2580.143
L2-SP–Anchor0.4570.2190.3160.4240.1570.2580.2420.1810.2090.3810.2650.318
EWC–Fisher anchor0.2750.2000.2350.2040.1120.1510.1050.1600.1300.2550.2580.257
GeoReplayGeometryAligned0.0960.1970.1370.080<u>0.098</u>0.0890.0750.1700.1130.0760.2530.139
GeoTransolver
FT (no replay)––0.1990.2050.2020.2360.0810.1380.226<u>0.148</u>0.1820.2230.2440.233
Plain replayUniformPlain0.0560.208<u>0.108</u>0.0540.0870.0690.0550.152<u>0.092</u>0.0550.2400.115
Weighting onlyGeometryPlain0.0560.2050.1070.0540.0860.0690.0550.1450.0900.0550.2370.115
Alignment onlyUniformAligned<u>0.059</u>0.2050.110<u>0.056</u><u>0.081</u>0.0680.0560.1550.0930.0580.2360.117
A-GEMUniformProjection0.0630.2070.1140.0580.0800.0680.0600.1560.0970.0600.2420.121
L2-SP–Anchor0.2530.2270.2400.3530.1430.2250.1880.1730.1800.2580.2480.253
EWC–Fisher anchor0.0980.2090.1430.0910.0940.0930.0590.1630.0980.0890.2410.146
GeoReplayGeometryAligned0.0600.2060.1110.0570.0820.0680.0580.1490.0930.0580.2370.117

Table 3.

Fine-tuning alone often loses to training from scratch. Without replay, fine-tuning has a lower pressure or wall shear stress error than training from scratch in only 17 of 40 comparisons at β=10\beta= 10 and a higher error in 23 (Appendix § 11). Which arm wins depends jointly on the surrogate, target dataset, and quantity. Fine-tuned GeoTransolver has a 41% higher pressure error than training from scratch on WindsorML yet stays within 2% of it on DrivAerNet++, and on SHIFT-SUV fine-tuned DoMINO lowers both surface-field errors while nearly tripling the drag error. Figure 5 in Appendix § 14 shows that a large reduction of zero-shot error can coexist with negative transfer value.

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\beta= 80 (β=79\beta= 79 for SHIFT-SUV). Bold marks the lowest relative L2L_2 error among SC, FT, and GR, and dashed ellipses mark the same surface region on their maps.

Distance to the closest training geometry does not determine transfer value. DrivAerNet++ and SHIFT-SUV lie closer to the DrivAerML training geometries than AhmedML and WindsorML do (Figure 2, Appendix Table 5), and Appendix § 8 (Figure 4) compares test-set coverage by the target training split and by DrivAerML. Nevertheless, fine-tuning has a lower pressure, wall shear stress, or drag error than training from scratch in only 5 of 15 surrogate and quantity comparisons on SHIFT-SUV, as rarely as on the more distant AhmedML (5 of 15) and WindsorML (4 of 15), against 9 of 15 on DrivAerNet++. Neither geometric distance nor zero-shot error determines transfer value, so geometric coverage alone cannot decide whether a pretrained surrogate remains useful after fine-tuning. The targets also differ from DrivAerML in CFD solver, turbulence model, scale, and Reynolds number (Table 1), which geometric distance does not capture.

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.

Figure 4. Minimum Chamfer distance from each held-out target geometry to the target training set (blue) or DrivAerML pretraining set (orange). Within each target, both distributions contain the same test geometries; boxes show the median and interquartile range, and whiskers extend to 1.5 times the interquartile range.

Discussion

Our proposed GeoReplay addresses the failure mode exposed by our transfer learning analysis. Fine-tuning on a few target simulations overwrites much of the source accuracy of the pretrained surrogate, and replaying source simulations during adaptation prevents this without a measurable loss of target accuracy.

Several limitations bound these conclusions. We measure variation over three training seeds at β=10\beta= 10 on one split per target, so small differences at other budgets require caution. Our geometry measures omit CFD solver, turbulence, operating-condition, mesh, and label differences that may affect transfer.

Conclusion

With few target simulations, fine-tuning a pretrained surrogate beats training from scratch in fewer than half of the comparisons and erodes its source accuracy. GeoReplay lowers the source error of fine-tuning in 77 of 80 settings, by a median factor of 2.8, while matching its target accuracy. Because geometric distance does not determine transfer value, evaluations should report both the from-scratch comparison and the source error. All released surrogates learn from DrivAerML, which is why we study this setting; pretraining across datasets is a natural next step, and Appendix H reports a first comparison.

AI use statement

We used generative AI tools to assist with programming, literature search, and editing the manuscript text. The authors reviewed and tested all AI-assisted code, verified all cited sources, and checked all reported results and text. We take responsibility for the final content of this work, including text, claims, and artifacts produced with the aid of generative AI.

Geometry Analysis

As a within-source calibration, we compare the official 48-geometry DrivAerML validation split with the same 436 pretraining geometries. Its median closest-training distance is 0.0183 body lengths, with an interquartile range of [0.0179, 0.0189].

We sample 1,024 area-weighted surface points per geometry and compute symmetric Chamfer distance by averaging the target-to-source and source-to-target nearest-point distances. We reconcile dataset axes and apply uniform body-length normalization before sampling, which produces a dimensionless distance while retaining differences in vehicle proportions. For Figure 2, we compare every usable geometry from the four target datasets with all 436 DrivAerML training geometries. For the within-dataset analysis, we form all unordered pairs within the 436 DrivAerML, 500 DrivAerNet++, 99 SHIFT-SUV and 350 WindsorML geometries. Table 5 retains the closest DrivAerML geometry for every usable target geometry, including the 499-geometry AhmedML extension. Figure 4 instead uses the ten fixed test geometries per complete target and compares their nearest DrivAerML and target-training geometries. As robustness checks, we also evaluate sliced Wasserstein distance, using 64 projection directions and 64 quantiles per direction, and a robust-standardized descriptor containing width-to-length and height-to-length ratios, vertical quantiles, and coarse longitudinal body profiles.

Target datasetNNMedian [IQR] ↓\downarrow95th pct. ↓\downarrow
DrivAerNet++5000.020[0.020, 0.020]0.021
SHIFT-SUV990.022[0.022, 0.022]0.022
AhmedML4990.036[0.031, 0.043]0.061
WindsorML3500.034[0.030, 0.043]0.052

Table 5. Nearest-source geometry distance for all usable geometries in four target datasets, measured against the closest of 436 DrivAerML pretraining geometries. Values are in body lengths; lower is closer. Bold/underlined mark the lowest/second-lowest.

Variance Across Different Seeds

TTErrorMethodTSol ↓\downarrowGeoT ↓\downarrowDoM ↓\downarrowXMGN ↓\downarrowGINO ↓\downarrow
DrivAerNet++eSe_SZS0.042±\pm0.0020.037±\pm0.0020.186±\pm0.0090.278±\pm0.0090.268±\pm0.003
DrivAerNet++SC0.736±\pm0.0050.764±\pm0.0020.830±\pm0.0040.719±\pm0.0030.777±\pm0.017
DrivAerNet++FT0.350±\pm0.0020.199±\pm0.0050.461±\pm0.0060.499±\pm0.0100.555±\pm0.005
DrivAerNet++GR0.096±\pm0.0040.060±\pm0.0000.135±\pm0.0040.392±\pm0.0090.360±\pm0.007
DrivAerNet++sGR0.290±\pm0.0030.063±\pm0.0010.232±\pm0.0060.371±\pm0.0080.432±\pm0.007
DrivAerNet++eTe_TZS1.020±\pm0.0301.271±\pm0.0370.794±\pm0.0421.194±\pm0.0550.865±\pm0.040
DrivAerNet++SC0.215±\pm0.0030.202±\pm0.0050.242±\pm0.0140.234±\pm0.0180.283±\pm0.019
DrivAerNet++FT0.197±\pm0.0000.205±\pm0.0020.214±\pm0.0130.226±\pm0.0130.283±\pm0.012
DrivAerNet++GR0.197±\pm0.0010.206±\pm0.0020.215±\pm0.0130.221±\pm0.0140.286±\pm0.012
DrivAerNet++sGR0.197±\pm0.0100.208±\pm0.0130.214±\pm0.0130.221±\pm0.0140.284±\pm0.012
DrivAerNet++eSeT\sqrt{e_Se_T}ZS0.207±\pm0.0060.217±\pm0.0060.384±\pm0.0140.576±\pm0.0160.481±\pm0.012
DrivAerNet++SC0.398±\pm0.0030.393±\pm0.0050.448±\pm0.0130.410±\pm0.0160.469±\pm0.016
DrivAerNet++FT0.263±\pm0.0010.202±\pm0.0010.314±\pm0.0100.336±\pm0.0110.397±\pm0.008
DrivAerNet++GR0.137±\pm0.0030.111±\pm0.0010.170±\pm0.0060.294±\pm0.0100.321±\pm0.007
DrivAerNet++sGR0.239±\pm0.0060.114±\pm0.0040.223±\pm0.0070.286±\pm0.0100.350±\pm0.008
SHIFT-SUVeSe_SZS0.042±\pm0.0020.037±\pm0.0020.186±\pm0.0090.278±\pm0.0090.268±\pm0.003
SHIFT-SUVSC0.722±\pm0.0120.739±\pm0.0051.261±\pm0.0300.814±\pm0.0230.720±\pm0.011
SHIFT-SUVFT0.414±\pm0.0020.236±\pm0.0050.485±\pm0.0050.714±\pm0.0220.972±\pm0.022
SHIFT-SUVGR0.080±\pm0.0010.057±\pm0.0000.134±\pm0.0040.445±\pm0.0140.428±\pm0.010
SHIFT-SUVsGR0.290±\pm0.0040.057±\pm0.0010.167±\pm0.0050.396±\pm0.0120.509±\pm0.013
SHIFT-SUVeTe_TZS0.733±\pm0.0050.845±\pm0.0050.687±\pm0.0031.110±\pm0.0150.582±\pm0.026
SHIFT-SUVSC0.103±\pm0.0030.067±\pm0.0010.134±\pm0.0070.103±\pm0.0090.221±\pm0.012
SHIFT-SUVFT0.097±\pm0.0000.081±\pm0.0000.119±\pm0.0070.117±\pm0.0090.239±\pm0.007
SHIFT-SUVGR0.098±\pm0.0000.082±\pm0.0010.124±\pm0.0070.111±\pm0.0090.239±\pm0.007
SHIFT-SUVsGR0.098±\pm0.0090.082±\pm0.0090.120±\pm0.0070.111±\pm0.0090.239±\pm0.007
SHIFT-SUVeSeT\sqrt{e_Se_T}ZS0.175±\pm0.0040.177±\pm0.0040.358±\pm0.0080.556±\pm0.0090.395±\pm0.009
SHIFT-SUVSC0.272±\pm0.0060.223±\pm0.0020.411±\pm0.0110.289±\pm0.0140.399±\pm0.011
SHIFT-SUVFT0.200±\pm0.0010.138±\pm0.0020.240±\pm0.0070.289±\pm0.0110.482±\pm0.009
SHIFT-SUVGR0.089±\pm0.0000.068±\pm0.0000.129±\pm0.0040.222±\pm0.0090.320±\pm0.006
SHIFT-SUVsGR0.168±\pm0.0080.069±\pm0.0040.141±\pm0.0050.210±\pm0.0090.349±\pm0.007
AhmedMLeSe_SZS0.042±\pm0.0020.037±\pm0.0020.186±\pm0.0090.278±\pm0.0090.268±\pm0.003
AhmedMLSC0.824±\pm0.0290.809±\pm0.0081.306±\pm0.0320.941±\pm0.0111.707±\pm0.040
AhmedMLFT0.272±\pm0.0220.226±\pm0.0200.286±\pm0.0070.396±\pm0.0080.579±\pm0.009
AhmedMLGR0.075±\pm0.0000.058±\pm0.0000.146±\pm0.0050.332±\pm0.0060.270±\pm0.004
AhmedMLsGR0.211±\pm0.0030.066±\pm0.0020.184±\pm0.0070.352±\pm0.0070.277±\pm0.004
AhmedMLeTe_TZS1.041±\pm0.0350.988±\pm0.0331.242±\pm0.0542.036±\pm0.0681.043±\pm0.045
AhmedMLSC0.136±\pm0.0050.124±\pm0.0030.243±\pm0.0510.329±\pm0.0490.590±\pm0.085
AhmedMLFT0.165±\pm0.0140.148±\pm0.0090.256±\pm0.0500.292±\pm0.0510.486±\pm0.097
AhmedMLGR0.170±\pm0.0080.149±\pm0.0040.253±\pm0.0480.288±\pm0.0520.474±\pm0.094
AhmedMLsGR0.165±\pm0.0520.145±\pm0.0560.254±\pm0.0480.297±\pm0.0470.470±\pm0.092
AhmedMLeSeT\sqrt{e_Se_T}ZS0.209±\pm0.0060.192±\pm0.0050.481±\pm0.0150.753±\pm0.0170.528±\pm0.012
AhmedMLSC0.335±\pm0.0070.317±\pm0.0060.563±\pm0.0610.557±\pm0.0411.003±\pm0.073
AhmedMLFT0.212±\pm0.0030.182±\pm0.0040.271±\pm0.0270.340±\pm0.0300.530±\pm0.054
AhmedMLGR0.113±\pm0.0020.093±\pm0.0020.192±\pm0.0190.309±\pm0.0280.358±\pm0.035
AhmedMLsGR0.187±\pm0.0290.098±\pm0.0190.216±\pm0.0210.323±\pm0.0260.361±\pm0.035
WindsorMLeSe_SZS0.042±\pm0.0020.037±\pm0.0020.186±\pm0.0090.278±\pm0.0090.268±\pm0.003
WindsorMLSC1.227±\pm0.2500.805±\pm0.0171.808±\pm0.0490.870±\pm0.0080.901±\pm0.008
WindsorMLFT0.333±\pm0.0320.223±\pm0.0330.296±\pm0.0060.451±\pm0.0080.867±\pm0.012
WindsorMLGR0.076±\pm0.0010.058±\pm0.0010.144±\pm0.0060.348±\pm0.0090.271±\pm0.004
WindsorMLsGR0.310±\pm0.0070.061±\pm0.0010.169±\pm0.0070.377±\pm0.0120.280±\pm0.004
WindsorMLeTe_TZS0.794±\pm0.0170.744±\pm0.0111.519±\pm0.0280.900±\pm0.0120.790±\pm0.013
WindsorMLSC0.427±\pm0.1130.173±\pm0.0050.279±\pm0.0180.329±\pm0.0200.470±\pm0.024
WindsorMLFT0.255±\pm0.0020.244±\pm0.0050.295±\pm0.0170.322±\pm0.0210.487±\pm0.026
WindsorMLGR0.253±\pm0.0010.237±\pm0.0070.291±\pm0.0190.324±\pm0.0210.484±\pm0.026
WindsorMLsGR0.256±\pm0.0170.237±\pm0.0170.294±\pm0.0170.323±\pm0.0210.484±\pm0.026
WindsorMLeSeT\sqrt{e_Se_T}ZS0.182±\pm0.0050.166±\pm0.0040.532±\pm0.0130.500±\pm0.0080.460±\pm0.005
WindsorMLSC0.724±\pm0.1690.373±\pm0.0020.710±\pm0.0240.535±\pm0.0160.651±\pm0.017
WindsorMLFT0.292±\pm0.0130.233±\pm0.0190.295±\pm0.0090.381±\pm0.0130.650±\pm0.018
WindsorMLGR0.139±\pm0.0010.117±\pm0.0020.205±\pm0.0080.335±\pm0.0120.362±\pm0.010
WindsorMLsGR0.282±\pm0.0100.120±\pm0.0050.223±\pm0.0080.349±\pm0.0130.368±\pm0.010

Table 6. Table 2 transposed, with the standard deviation over three training seeds. Columns are surrogates, and rows give the source error eSe_S, the target error eTe_T, and their geometric mean (shaded) for each method and target dataset TT. 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.

GeoReplay Budget Sweep

In Table 7, we compare GeoReplay and sparse GeoReplay with ordinary fine-tuning and training from scratch across five surrogates and target training budgets β∈{1,10,40,80}\beta\in\{1, 10, 40, 80\} (79 for SHIFT-SUV at the largest budget). We lay out Table 7 (pressure), 8 (wall shear stress), 9 (drag), and 10 (lift) like Table 2, grouping rows by dataset, budget, and surrogate and reporting the DrivAerML source error, the target error, and their geometric mean for zero-shot prediction and the four adapted methods. We report SAV in Appendix D. We select checkpoints using target validation errors and evaluate them on the fixed held-out test simulations.

Table 7: Pressure errors of GeoReplay across target datasets TT, target training budgets β\beta and surrogates, laid out as Table 2. The blocks give the source error eSe_S of MSTM_S^T on the 48 designs of the official DrivAerML validation split, the target error eTe_T of MTTM_T^T 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\beta. In each block, bold marks the lowest value in each row.

TTβ\betaSurrogateSource error eSe_S ZS ↓SC ↓FT ↓GR ↓sGR ↓Target error eTe_T ZS ↓SC ↓FT ↓GR ↓sGR ↓Geometric mean eSeT\sqrt{e_S e_T} ZS ↓SC ↓FT ↓GR ↓sGR ↓
DrivAerNet++1TSol0.0420.7890.3130.2310.3061.0200.5040.3470.3470.3480.2070.6300.3290.2830.326
GeoT0.0370.7520.1660.0920.1191.2710.3810.3750.3760.3750.2170.5350.2490.1860.211
DoM0.1860.8040.4220.3480.3960.7940.4030.3290.3290.3290.3840.5690.3730.3390.361
XMGN0.2780.7260.4790.4630.4491.1940.3930.3730.3680.3670.5760.5340.4230.4120.406
GINO0.2680.7640.5110.4810.4980.8650.5090.4040.4040.4040.4810.6240.4540.4410.449
10TSol0.0420.7360.3500.0960.2901.0200.2150.1970.1970.1970.2070.3980.2630.1370.239
GeoT0.0370.7640.1990.0600.0631.2710.2020.2050.2060.2080.2170.3930.2020.1110.114
DoM0.1860.8300.4610.1350.2320.7940.2420.2140.2150.2140.3840.4480.3140.1700.223
XMGN0.2780.7190.4990.3920.3711.1940.2340.2260.2210.2210.5760.4100.3360.2940.286
GINO0.2680.7770.5550.3600.4320.8650.2830.2830.2860.2840.4810.4690.3970.3210.350
40TSol0.0420.7240.3830.0820.2411.0200.1800.1700.1700.1710.2070.3610.2550.1180.203
GeoT0.0370.7440.2370.0580.0601.2710.1720.1710.1730.1700.2170.3580.2010.1010.101
DoM0.1860.7440.5310.1240.1490.7940.1960.2110.2180.20.3840.3810.3350.1650.183
XMGN0.2780.7360.5270.3390.3291.1940.2030.1970.1930.1930.5760.3870.3220.2560.252
GINO0.2680.7260.5710.2880.3420.8650.2470.2490.2520.2500.4810.4240.3770.2700.292
80TSol0.0420.7460.3940.0810.2271.0200.1750.1600.1610.1630.2070.3610.2510.1150.193
GeoT0.0370.7500.2210.0570.0581.2710.1720.1620.1630.1640.2170.3600.1890.0960.097
DoM0.1860.7400.5630.1190.1390.7940.1790.1890.1790.1830.3840.3640.3260.1460.159
XMGN0.2780.6960.5250.3180.2881.1940.1890.1880.1820.1820.5760.3630.3140.2410.229
GINO0.2680.7420.5690.2790.3090.8650.2280.2310.2360.2330.4810.4110.3630.2570.268
SHIFT-SUV1TSol0.0420.8110.2780.1990.2680.7330.4100.3330.3330.3330.1750.5770.3040.2580.299
GeoT0.0370.8200.1970.1140.1100.8450.2440.3410.3490.3080.1770.4470.2600.1990.184
DoM0.1860.9430.4320.3650.3930.6870.3280.2290.2290.2290.3580.5560.3150.2890.300
XMGN0.2781.1460.5950.6320.6351.1100.2440.2670.2440.2440.5560.5290.3990.3930.394
GINO0.2680.6520.4760.4810.4770.5820.3900.3400.3400.3400.3950.5040.4020.4040.402
10TSol0.0420.7220.4140.0800.2900.7330.1030.0970.0980.0980.1750.2720.2000.0890.168
GeoT0.0370.7390.2360.0570.0570.8450.0670.0810.0820.0820.1770.2230.1380.0680.069
DoM0.1861.2610.4850.1340.1670.6870.1340.1190.1240.1200.3580.4110.2400.1290.141
XMGN0.2780.8140.7140.4450.3961.1100.1030.1170.1110.1110.5560.2890.2890.2220.210
GINO0.2680.7200.9720.4280.5090.5820.2210.2390.2390.2390.3950.3990.4820.3200.349
40TSol0.0420.7340.4940.0750.2530.7330.0560.0680.0700.0690.1750.2030.1830.0720.132
GeoT0.0370.7330.3950.0550.0560.8450.0520.0580.0570.0580.1770.1950.1510.0560.057
DoM0.1860.9260.6010.1250.1350.6870.0810.0910.1140.0960.3580.2730.2340.1190.114
XMGN0.2780.7570.7030.3880.2991.1100.0860.0920.0980.0900.5560.2560.2540.1950.164
GINO0.2680.7501.2440.2870.3190.5820.1790.2090.2130.2110.3950.3670.5090.2470.259
79TSol0.0420.7100.5630.0740.2560.7330.0500.0620.0640.0630.1750.1890.1860.0690.127
GeoT0.0370.7190.4720.0550.0550.8450.0470.0540.0530.0540.1770.1840.1600.0540.055
DoM0.1860.8580.6660.1210.1320.6870.0630.0780.1020.0910.3580.2320.2280.1110.110
XMGN0.2780.8330.7280.3630.2861.1100.0870.0950.0930.0880.5560.2700.2640.1840.159
GINO0.2680.7621.4200.2800.3000.5820.1670.1930.1990.1960.3950.3570.5230.2360.242
AhmedML1TSol0.0420.9780.1960.0730.1801.0410.4080.3290.3300.3300.2090.6320.2540.1550.244
GeoT0.0370.8830.1780.0830.0840.9880.2870.3380.3350.3360.1920.5040.2450.1670.168
DoM0.1860.9410.2670.2310.2261.2420.4200.4740.4720.4730.4810.6290.3560.3310.327
XMGN0.2780.9100.3750.3630.3762.0360.8380.5470.5470.5460.7530.8730.4530.4460.453
GINO0.2680.8120.5110.4790.4881.0430.8900.6910.6900.6910.5280.8500.5940.5750.580
10TSol0.0420.8240.2720.0750.2111.0410.1360.1650.1700.1650.2090.3350.2120.1130.187
GeoT0.0370.8090.2260.0580.0660.9880.1240.1480.1490.1450.1920.3170.1820.0930.098
DoM0.1861.3060.2860.1460.1841.2420.2430.2560.2530.2540.4810.5630.2710.1920.216
XMGN0.2780.9410.3960.3320.3522.0360.3290.2920.2880.2970.7530.5570.3400.3090.323
GINO0.2681.7070.5790.2700.2771.0430.5900.4860.4740.4700.5281.0030.5300.3580.361
40TSol0.0420.8680.5190.0720.2531.0410.0900.0840.0900.0910.2090.2800.2090.0810.152
GeoT0.0370.7840.4180.0560.0570.9880.0710.0830.0740.0840.1920.2350.1860.0640.069
DoM0.1861.0090.3630.1220.1331.2420.2430.2290.2380.2410.4810.4950.2880.1700.183
XMGN0.2780.8670.5490.3140.3712.0360.2390.2230.2320.2290.7530.4550.3500.2700.291

Table 7.

Continued on next page.

TTβ\betaSurrogate\multicolumn{5}{c}{Source error eSe_S}\multicolumn{5}{c}{Target error eTe_T}\multicolumn{5}{c}{Geometric mean eSeT\sqrt{e_S e_T}}
ZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrow
GINO0.2681.0850.6340.2690.2721.0430.3730.3780.3870.3870.5280.6360.4890.3230.325
80TSol0.0420.8230.5880.0730.3011.0410.0900.0920.1000.0850.2090.2720.2320.0860.160
GeoT0.0370.7560.5430.0560.0550.9880.0790.0850.0910.0840.1920.2440.2150.0710.068
DoM0.1861.0100.3850.1170.1271.2420.2570.1930.1880.1940.4810.5100.2720.1490.157
XMGN0.2780.8550.6360.2840.2722.0360.2050.1900.1860.1860.7530.4190.3470.2300.225
GINO0.2680.9130.9310.2700.2751.0430.3150.3500.3530.3520.5280.5360.5710.3090.311
WindsorML1TSol0.0420.9810.2640.1050.2540.7940.5090.4690.4700.4700.1820.7070.3510.2220.346
GeoT0.0371.0670.1820.1060.1710.7440.5060.4570.4570.4580.1660.7350.2880.2200.280
DoM0.1861.0830.2370.2120.2111.5190.4900.5080.5060.5070.5320.7280.3470.3280.327
XMGN0.2780.9200.3760.3980.4130.9000.6190.5240.5110.5100.5000.7550.4440.4510.459
GINO0.2680.8870.5130.4230.4590.7900.6600.5940.5940.5950.4600.7650.5520.5010.522
10TSol0.0421.2270.3330.0760.3100.7940.4270.2550.2530.2560.1820.7240.2920.1390.282
GeoT0.0370.8050.2230.0580.0610.7440.1730.2440.2370.2370.1660.3730.2330.1170.120
DoM0.1861.8080.2960.1440.1691.5190.2790.2950.2910.2940.5320.7100.2950.2050.223
XMGN0.2780.8700.4510.3480.3770.9000.3290.3220.3240.3230.5000.5350.3810.3350.349
GINO0.2680.9010.8670.2710.2800.7900.4700.4870.4840.4840.4600.6510.6500.3620.368
40TSol0.0420.8130.6070.0720.3650.7940.1190.1440.1440.1440.1820.3110.2960.1020.230
GeoT0.0370.8490.3670.0540.0560.7440.1120.1180.1160.1160.1660.3080.2080.0790.081
DoM0.1861.7800.3890.1290.1431.5190.1960.2100.2120.2090.5320.5910.2860.1650.173
XMGN0.2780.8790.5420.3080.2850.9000.2620.2360.2360.2320.5000.4790.3570.2690.257
GINO0.2681.1221.2090.2690.2750.7900.3860.4410.4430.4420.4600.6580.7300.3450.348
80TSol0.0420.8300.6460.0720.3500.7940.1040.1250.1250.1250.1820.2940.2840.0950.209
GeoT0.0370.8340.5630.0550.0550.7440.0970.1050.1060.1050.1660.2850.2430.0760.076
DoM0.1862.9510.5740.1190.1321.5190.1430.1610.1660.1610.5320.6490.3040.1410.146
XMGN0.2780.9070.5850.2830.2850.9000.2380.1740.1720.1740.5000.4640.3190.2210.223
GINO0.2681.3812.0250.2690.2760.7900.3320.4020.4130.4100.4600.6770.9020.3330.336

Table 7 (PDF p. 15).

TTβ\betaSurrogate\multicolumn{5}{c}{Source error eSe_S}\multicolumn{5}{c}{Target error eTe_T}\multicolumn{5}{c}{Geometric mean eSeT\sqrt{e_S e_T}}
ZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrow
DrivAerNet++1TSol0.0790.8990.4510.3130.4491.1550.7090.4670.4670.4680.3020.7980.4590.3820.458
GeoT0.0700.9480.2830.1520.1941.0190.4770.4920.4920.4930.2670.6730.3730.2740.310
DoM0.2801.2480.5500.4720.5251.0910.5120.4450.4460.4450.5530.7990.4950.4590.483
XMGN0.4001.1160.6470.6320.6011.1930.4730.4580.4550.4550.6910.7270.5440.5360.523
GINO0.4820.9780.9430.8900.9181.2080.6270.5410.5410.5410.7630.7830.7140.6940.705
10TSol0.0790.9160.5000.1600.4481.1550.3150.3000.3000.2990.3020.5370.3870.2190.366
GeoT0.0700.9230.3180.1140.1191.0190.3150.3070.3060.3100.2670.5390.3130.1870.192
DoM0.2801.1300.6420.2130.3611.0910.3420.3170.3190.3180.5530.6210.4510.2610.339
XMGN0.4000.9800.6980.4800.4551.1930.3350.3380.3310.3310.6910.5730.4850.3990.388
GINO0.4821.0801.1040.5780.7331.2080.3800.3860.3890.3870.7630.6400.6530.4740.532
40TSol0.0790.9290.5370.1460.3711.1550.2790.2670.2680.2650.3020.5090.3780.1980.314
GeoT0.0700.9210.3610.1090.1121.0190.2750.2710.2710.2720.2670.5030.3130.1720.174
DoM0.2801.1040.6450.1920.2341.0910.3120.3110.3190.3280.5530.5870.4480.2470.277
XMGN0.4000.9980.7420.4270.4251.1930.3020.3040.2980.2980.6910.5490.4750.3570.356
GINO0.4821.0861.1660.5060.5641.2080.3390.3450.3490.3460.7630.6070.6340.4200.442
80TSol0.0790.9040.5440.1450.3581.1550.2680.2530.2530.2540.3020.4930.3710.1920.301
GeoT0.0700.9210.3280.1060.1081.0190.2650.2580.2600.2570.2670.4940.2910.1660.167
DoM0.2801.0480.6670.1850.2161.0910.2710.2980.2800.2870.5530.5330.4460.2280.249
XMGN0.4000.9870.7430.4130.3811.1930.2940.2950.2900.2900.6910.5390.4680.3460.332
GINO0.4821.0691.1650.4970.5301.2080.3230.3240.3280.3250.7630.5870.6140.4040.415
SHIFT-SUV1TSol0.0791.0120.4370.3130.4261.4220.4650.4300.4300.4300.3350.6860.4340.3670.428
GeoT0.0701.0260.3200.1700.1571.2640.3150.4400.4390.3940.2970.5680.3750.2740.249
DoM0.2801.0850.5640.4880.5221.2640.3970.3480.3490.3490.5950.6570.4430.4130.427
XMGN0.4001.0850.6540.6550.6301.9370.3190.3500.3310.3310.8800.5880.4780.4650.457
GINO0.4821.0120.9130.8880.8991.7310.4810.5040.5050.5050.9130.6980.6780.6700.674
10TSol0.0791.0080.7100.1430.4941.4220.1720.1790.1820.1810.3350.4170.3570.1620.299
GeoT0.0701.0040.4060.1070.1071.2640.1000.1330.1360.1390.2970.3160.2320.1200.122
DoM0.2801.1950.6240.1950.2331.2640.2540.2060.2190.2130.5950.5510.3580.2070.223
XMGN0.4001.0360.7540.5280.4851.9370.1670.1980.1900.1900.8800.4150.3860.3160.304
GINO0.4821.0071.0270.6960.8071.7310.3650.3880.3890.3880.9130.6060.6310.5210.560
40TSol0.0791.0130.8400.1350.4231.4220.0910.1220.1240.1230.3350.3040.3200.1300.228
GeoT0.0701.0070.7120.1020.1031.2640.0790.0950.0940.0960.2970.2820.2600.0980.099
DoM0.2801.1960.6820.1850.1941.2640.1470.1470.2030.1610.5950.4190.3170.1940.176
TTβ\betaSurrogate\multicolumn{5}{c}{Source error eSe_S}\multicolumn{5}{c}{Target error eTe_T}\multicolumn{5}{c}{Geometric mean eSeT\sqrt{e_S e_T}}
ZS ↓SC ↓FT ↓GR ↓sGR ↓ZS ↓SC ↓FT ↓GR ↓sGR ↓ZS ↓SC ↓FT ↓GR ↓sGR ↓
XMGN0.4001.0320.8200.4760.3991.9370.1540.1630.1730.1620.8800.3980.3660.2870.254
GINO0.4821.0051.0550.5040.5461.7310.3210.3540.3600.3560.9130.5680.6110.4260.441
79TSol0.0791.0230.8930.1330.4211.4220.0780.1060.1120.1080.3350.2820.3070.1220.214
GeoT0.0701.0070.8010.1020.1021.2640.0720.0870.0840.0870.2970.2690.2630.0930.094
DoM0.2801.1190.7420.1800.1911.2640.1020.1240.1810.1550.5950.3390.3040.1800.172
XMGN0.4001.0340.7830.4590.3851.9370.1550.1710.1660.1590.8800.4000.3660.2760.248
GINO0.4821.0161.0450.4940.5221.7310.3060.3400.3470.3430.9130.5580.5960.4140.423
AhmedML1TSol0.0791.0340.3570.1320.3460.8760.3110.2440.2470.2460.2630.5670.2950.1810.292
GeoT0.0700.9960.2960.1500.1500.8000.2210.2340.2330.2350.2360.4700.2630.1870.188
DoM0.2801.0960.3650.3240.3290.7930.3000.3400.3390.3410.4720.5730.3520.3320.335
XMGN0.4001.2330.4900.4690.4920.9100.5440.3920.3900.3890.6030.8190.4380.4280.438
GINO0.4820.9200.7010.6610.6640.7750.6550.6300.6300.6290.6110.7760.6650.6450.647
10TSol0.0790.9280.4800.1390.3980.8760.1160.1330.1350.1340.2630.3280.2520.1370.231
GeoT0.0700.8850.3610.1100.1210.8000.1080.1150.1170.1150.2360.3090.2030.1130.118
DoM0.2801.3210.3770.2190.2610.7930.1720.1730.1720.1710.4720.4760.2550.1940.212
XMGN0.4001.1470.5150.4160.4340.9100.2340.2080.2040.2090.6030.5180.3270.2910.301
GINO0.4821.3660.7830.4830.4890.7750.4430.3690.3740.3720.6110.7770.5380.4250.427
40TSol0.0791.0200.6700.1320.3970.8760.0880.0860.0870.0910.2630.2990.2400.1070.190
GeoT0.0700.9180.5980.1030.1070.8000.0760.0790.0760.0840.2360.2640.2170.0890.094
DoM0.2801.1450.4220.1860.2020.7930.1570.1590.1610.1610.4720.4250.2590.1730.180
XMGN0.4000.9760.6980.3980.4710.9100.1580.1530.1600.1580.6030.3920.3260.2520.273
GINO0.4821.1750.8840.4840.4860.7750.3020.3040.3100.3090.6110.5960.5190.3870.388
80TSol0.0790.9880.7600.1350.4650.8760.0820.0910.0960.0870.2630.2840.2630.1140.201
GeoT0.0700.8990.7340.1040.1040.8000.0800.0870.0910.0840.2360.2670.2530.0970.093
DoM0.2801.1580.5120.1770.1920.7930.1930.1400.1340.1410.4720.4730.2680.1540.165
XMGN0.4000.9170.7580.3780.3700.9100.1380.1250.1230.1240.6030.3560.3070.2160.215
GINO0.4821.1931.0360.4830.4880.7750.2670.2800.2830.2820.6110.5640.5390.3700.371
WindsorML1TSol0.0791.0760.4260.2010.4080.8770.3470.3400.3410.3410.2630.6110.3810.2620.373
GeoT0.0701.1270.3170.1790.2950.7930.3350.3120.3120.3110.2350.6150.3140.2360.303
DoM0.2801.5010.3670.2830.3011.3440.3570.3270.3290.3280.6140.7320.3470.3050.315
XMGN0.4001.0360.5610.5610.5680.8890.4320.3430.3420.3410.5960.6690.4380.4380.440
GINO0.4821.1520.7160.5900.6560.8090.4740.4440.4430.4440.6250.7390.5640.5110.540
10TSol0.0791.1980.5050.1390.4670.8770.2960.1930.1890.1890.2630.5950.3120.1620.297
GeoT0.0701.0190.3550.1100.1170.7930.1310.1750.1760.1760.2350.3660.2490.1390.143
DoM0.2801.6440.4220.2130.2421.3440.2040.2180.2150.2190.6140.5800.3030.2140.230
XMGN0.4000.9990.5990.4690.5200.8890.2370.2220.2280.2270.5960.4860.3650.3270.343
GINO0.4821.1261.0340.4850.4930.8090.3520.3630.3610.3620.6250.6290.6130.4180.422
40TSol0.0791.0370.8290.1330.5390.8770.1000.1170.1190.1180.2630.3220.3120.1260.252
GeoT0.0701.0160.5860.1020.1050.7930.0950.1030.1020.1020.2350.3110.2460.1020.104
DoM0.2801.6600.5070.1930.2131.3440.1490.1630.1660.1650.6140.4980.2870.1790.187
XMGN0.4001.0280.7210.4260.3980.8890.1820.1630.1640.1600.5960.4330.3430.2640.253
GINO0.4821.2481.3040.4820.4870.8090.2920.3270.3280.3270.6250.6040.6530.3980.399
80TSol0.0791.0330.8650.1310.5230.8770.0910.1050.1050.1060.2630.3070.3010.1170.235
GeoT0.0701.0500.8110.1020.1020.7930.0850.0930.0940.0940.2350.2980.2750.0980.098
DoM0.2802.3530.6270.1790.1991.3440.1100.1240.1310.1280.6140.5100.2790.1530.159
XMGN0.4001.0410.7940.4010.4150.8890.1640.1290.1290.1300.5960.4140.3210.2270.232
GINO0.4821.2532.6510.4820.4870.8090.2600.3060.3120.3110.6250.5710.9000.3880.389

Table 8. Wall shear stress errors of GeoReplay across target datasets TT, target training budgets β\beta and surrogates, laid out as Table 2. The blocks give the source errors of eSMe_S^M on the 48 designs of the official DrivAerML validation split, the target error of eTMe_T^M 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\beta. In each block, bold marks the lowest value in each row.

Table 9: Drag errors of GeoReplay across target datasets TT, target training budgets β\beta and surrogates, laid out as Table 2. The blocks give the source error eSe_S of MTTM_T^T on the 48 designs of the official DrivAerML validation split, the target error eTe_T of MTTM_T^T 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\beta. In each block, bold marks the lowest value in each row.

TTβ\betaSurrogate\multicolumn{5}{c}{Source error eSe_S}\multicolumn{5}{c}{Target error eTe_T}\multicolumn{5}{c}{Geometric mean eSeT\sqrt{e_S e_T}}
ZS ↓SC ↓FT ↓GR ↓sGR ↓ZS ↓SC ↓FT ↓GR ↓sGR ↓ZS ↓SC ↓FT ↓GR ↓sGR ↓
DrivAerNet++1TSol0.01070.55940.38400.40190.39020.24460.10880.12970.13030.13280.05120.24670.22320.22880.2276
GeoT0.00840.28820.03230.06470.08840.50010.11340.22480.22420.22140.06480.18080.08520.12050.1399
DoM0.01640.06400.35890.35930.36150.85290.16540.07320.07250.07280.11810.41950.16210.16140.1622
XMGN0.08930.96030.24210.24290.26150.35640.10550.10490.10440.10440.17840.43160.15980.15960.1653
GINO0.11770.60880.28050.28450.28220.21190.21780.15960.15940.15950.15790.36410.21160.21290.2121
10TSol0.01070.44720.46710.04350.47250.24460.04830.03700.03630.03270.05120.14700.13140.03980.1243
GeoT0.00840.37390.07830.00870.01060.50010.03980.02110.04120.03040.06480.12190.04060.01890.0180
DoM0.01640.20380.36220.02840.10820.85290.04980.03260.02890.02790.11810.10080.10870.02870.0549
XMGN0.08930.41160.34650.46390.46680.35640.08020.08310.06510.06460.17840.18170.16970.17370.1693
GINO0.11770.69640.39760.37030.40160.21190.10720.10730.10570.10840.15790.27330.20650.19780.2086
40TSol0.01070.59790.43000.03200.25030.24460.03020.02840.03590.03790.05120.13450.11060.03390.0974
GeoT0.00840.43380.13170.00980.01380.50010.03510.02650.03280.03740.06480.12340.05900.01790.0227
DoM0.01640.36320.38220.02220.03740.85290.03140.02370.02780.01990.11810.10680.09520.02490.0273
XMGN0.08930.52020.23960.54430.51050.35640.05030.06010.07410.06870.17840.16180.12000.20090.1873
TTβ\betaSurrogate\multicolumn{5}{c}{Source error eSe_S}\multicolumn{5}{c}{Target error eTe_T}\multicolumn{5}{c}{Geometric mean eSeT\sqrt{e_S e_T}}
ZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrow
GINO0.11770.69700.40530.19350.32970.21190.11560.07000.07090.07030.15790.28380.16840.11720.1523
80TSol0.01070.52290.49190.04230.17570.24460.03870.06540.02820.02840.05120.14230.17940.03450.0706
GeoT0.00840.42130.12010.01240.00870.50010.02790.01640.03370.01860.06480.10850.04440.02040.0127
DoM0.01640.41570.37290.02400.02900.85290.04420.02300.04720.02250.11810.13550.09260.03360.0255
XMGN0.08930.43120.20260.50080.42850.35640.06020.05920.04300.04350.17840.16110.10960.14680.1365
GINO0.11770.66260.46110.17240.25140.21190.05860.06520.06890.06310.15790.19710.17340.10900.1260
SHIFT-SUV1TSol0.01070.58710.05830.09190.06301.33870.04060.05410.05050.05330.11980.15430.05620.06810.0579
GeoT0.00840.96240.08430.05220.06081.99360.05620.24770.23900.13050.12950.23260.14450.11170.0891
DoM0.01640.56040.37240.34650.35880.81810.11490.06850.06960.06990.11570.25370.15970.15530.1584
XMGN0.08930.62490.59700.65070.63991.74310.08470.08710.09200.09330.39460.23010.22800.24460.2444
GINO0.11770.63420.15240.15360.15290.29230.26190.11980.12040.11970.18550.40760.13510.13600.1353
10TSol0.01070.90680.24530.01420.22641.33870.03620.01730.01710.01560.11980.18120.06520.01560.0594
GeoT0.00840.84610.18300.01020.01011.99360.00950.01820.01760.01750.12950.08940.05770.01340.0133
DoM0.01641.10150.43020.02320.03000.81810.01320.03940.03940.04110.11570.12070.13010.03020.0351
XMGN0.08930.11720.75800.39010.32601.74310.02470.02210.02410.02330.39460.05380.12930.09700.0871
GINO0.11770.42730.28840.27400.28850.29230.07430.08620.08850.08700.18550.17820.15770.15570.1585
40TSol0.01070.84720.22600.01230.10271.33870.00610.00740.00720.00810.11980.07200.04080.00940.0288
GeoT0.00840.88090.32120.00790.00861.99360.00940.00770.00490.00480.12950.09110.04960.00620.0064
DoM0.01641.54130.51140.02130.02510.81810.01180.03700.05200.03310.11570.13510.13750.03330.0288
XMGN0.08930.09840.71660.25320.11101.74310.01200.01200.01750.01100.39460.03440.09270.06660.0349
GINO0.11770.47340.40530.11780.09880.29230.04290.06700.06590.06380.18550.14250.16480.08810.0794
79TSol0.01070.70610.28520.01340.12501.33870.00570.00520.00740.00910.11980.06350.03860.01000.0337
GeoT0.00840.66700.47480.01060.00971.99360.00580.01080.00450.00470.12950.06240.07170.00700.0067
DoM0.01641.71200.55840.02200.02440.81810.01420.04280.00590.03470.11570.15600.15450.03600.0291
XMGN0.08930.21910.73620.17040.12641.74310.01800.01550.02240.01040.39460.06280.10680.06180.0363
GINO0.11770.36670.44050.11150.12781.74310.03650.05760.05960.05150.18550.11560.15920.08150.0811
AhmedML1TSol0.01071.36180.04110.01820.03020.84960.32020.20090.19630.19820.09550.66040.09080.05980.0774
GeoT0.00840.32560.17670.01610.01440.86710.27510.17070.17890.18200.08540.29930.05480.05360.0512
DoM0.01641.12030.02370.02420.02560.36630.60920.26970.26560.27000.07740.82610.07990.08010.0831
XMGN0.08931.83830.39420.40260.39122.15011.02830.63290.61420.61150.43830.43480.49950.49720.4891
GINO0.11770.94661.29261.24221.24911.98831.23060.53170.52180.52970.48371.07930.82900.80510.8134
10TSol0.01070.74620.03410.01430.04680.84960.11650.12860.13230.12530.09550.29480.06620.04350.0766
GeoT0.00840.49320.02850.00870.01090.86710.08270.13420.11270.12000.08540.20190.06180.03130.0362
DoM0.01640.69010.13740.01800.03330.36630.25370.28330.26120.26350.07740.41840.19730.06860.0937
XMGN0.08931.12430.44560.46770.47322.15010.21920.27030.25600.26490.43830.49650.34700.34600.3540
GINO0.11771.25751.06670.05080.08011.98830.73520.59710.53950.53360.48370.96150.79810.16550.2067
40TSol0.01070.35280.11100.01720.11170.84960.07570.05460.05510.05290.09550.16350.07780.03080.0763
GeoT0.00840.61780.07520.00860.00990.86710.05740.06030.05830.05790.08540.18830.06730.02230.0239
DoM0.01640.23070.13280.01800.01660.36630.24930.18330.21510.21200.07740.23980.15600.06220.0593
XMGN0.08930.11050.72950.49810.49602.15010.18510.16400.16080.13060.43830.14310.34590.28300.2545
GINO0.11771.05340.26440.07700.06601.98830.34680.31020.32810.32800.48370.60440.28640.15900.1471
80TSol0.01070.89690.10220.01630.10230.84960.05390.07100.07760.05970.09550.21980.08520.03560.0782
GeoT0.00840.11720.06680.01060.01000.86710.06860.06430.06070.05860.08540.08970.06560.02540.0242
DoM0.01640.50150.08420.02020.02190.36630.18370.11020.13300.11980.07740.30350.09630.05190.0512
XMGN0.08930.27350.98800.27010.21882.15010.16230.10150.11130.10960.43830.21070.31670.17340.1548
GINO0.11771.46140.46740.08150.06821.98830.26380.30420.31890.31860.48370.62090.37710.16120.1474
WindsorML1TSol0.01071.31760.17120.10880.15170.71860.41830.24930.24940.24520.08780.74240.20660.16480.1929
GeoT0.00841.08000.04500.03060.04230.99460.41650.28340.28430.28390.09140.86780.11290.09330.1096
DoM0.01641.38360.27760.05360.13210.71680.41580.29510.30110.29680.10830.75840.28620.12700.1980
XMGN0.08931.00930.47220.48440.47091.65470.28560.34670.34440.34450.38450.53690.40460.40840.4028
GINO0.11771.49150.50840.31100.42651.76600.61980.60790.59490.61080.45590.96150.55590.43010.5104
10TSol0.01071.50190.03120.01370.03540.71860.27610.16090.13950.20210.08780.64400.07090.04370.0846
GeoT0.00841.02540.03100.01140.00850.99460.12630.16420.14070.15640.09140.35990.07130.04000.0365
DoM0.01640.87140.35900.01730.02330.71680.15360.16510.16690.16690.10830.36580.24350.05370.0624
XMGN0.08931.38380.49890.37020.37151.65470.40520.34940.36340.37070.38450.74880.41750.36680.3711
GINO0.11770.92351.36410.05390.04791.76600.20810.38160.33870.34740.45590.43840.72150.13510.1290
40TSol0.01070.81280.24550.01390.11270.71860.06900.05780.05230.05430.08780.23690.11910.02700.0782
GeoT0.00840.98810.05900.00940.00910.99460.07590.05900.05700.06000.09140.27390.05900.02320.0233
DoM0.01640.51480.30970.01880.01900.71680.10940.14570.13580.14010.10830.23740.21240.05050.0516
XMGN0.08930.13470.27170.15250.16171.65470.34320.31610.32130.31300.38450.21500.29300.22130.2250
GINO0.11770.64850.09110.09780.07201.76600.14900.17820.22180.21880.45590.31090.12740.14730.1255
80TSol0.01071.48610.27750.01290.08340.71860.05880.04960.06480.04940.08780.29570.11730.02890.0642
GeoT0.00841.16880.12130.00910.00940.99460.04720.04970.05010.05190.09140.25300.07770.02130.0220
DoM0.01641.44190.08380.01910.02530.71680.07020.07710.09510.08520.10830.31820.08040.04260.0465
XMGN0.08930.18770.21210.18010.11551.65470.31890.27480.23800.28280.38450.24740.24140.20700.1776
GINO0.11770.25930.19710.10360.08321.76600.10670.13190.13320.14400.45590.16630.16120.11750.1094

Table 9. Table 9 continued

TTβ\betaSurrogateSource error eSe_S: ZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowTarget error eTe_T: ZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrowGeometric mean eSeT\sqrt{e_S e_T}: ZS ↓\downarrowSC ↓\downarrowFT ↓\downarrowGR ↓\downarrowsGR ↓\downarrow
DrivAerNet++1TSol0.2569.1882.3521.2532.3573.2851.7010.5000.4660.4830.9173.9541.0850.7641.067
GeoT0.2752.5271.1460.3260.2803.9022.3705.6935.2815.9351.0352.4472.5541.3131.289
DoM1.6048.7883.4593.7763.59414.9031.6604.3654.1554.3364.8893.8203.8863.9613.947
XMGN2.30017.55911.43911.38111.29857.4043.8432.8863.0282.94511.4908.2155.7465.8705.768
GINO0.9885.6970.9130.9220.9114.0652.2692.6802.6842.6902.0043.5951.5641.5731.565
10TSol0.25614.4223.9830.5363.3863.2853.2721.2780.9451.2440.9176.8692.2570.7122.052
GeoT0.2755.8911.7010.5570.3663.9022.0893.3492.5182.3331.0353.5082.3871.1840.924
DoM1.6049.9672.6040.7101.89614.9033.9292.7943.1323.1484.8896.2582.6971.4922.443
XMGN2.3008.26511.4588.7477.58657.4042.9312.6752.2102.24911.4904.9225.5374.3964.130
GINO0.9883.2523.5273.0783.8664.0651.1351.8731.7501.8492.0041.9212.5702.3212.674
40TSol0.2569.5342.8400.3413.1273.2852.9880.7960.9010.3490.9175.3381.5040.5541.044
GeoT0.2759.4041.5700.3860.4403.9021.2801.4092.2901.5251.0353.4701.4870.9410.819
DoM1.6042.4363.2760.5671.14114.9031.6341.6661.4861.6144.8891.9952.3360.9181.357
XMGN2.3006.03912.1273.9982.31357.4042.4761.7761.8501.87011.4903.8674.6412.7202.080
GINO0.9882.0904.0101.1372.2334.0651.3061.3191.2671.2112.0041.6522.3001.2001.644
80TSol0.2566.4552.2560.2702.3743.2852.6411.2561.0071.2010.9174.1291.6840.5221.689
GeoT0.2757.0393.4170.2800.2713.9022.3511.9702.0811.8751.0354.0682.5950.7630.713
DoM1.6042.0182.9070.5040.83514.9031.8701.5682.1391.5014.8891.9432.1351.0381.119
XMGN2.3005.07212.2821.9161.77857.4042.3221.8561.9401.92211.4903.4324.7741.9281.848
GINO0.9881.6615.1970.8571.4984.0651.8221.9811.8761.8962.0041.7393.2081.2681.685
SHIFT-SUV1TSol0.25639.0171.2322.5961.2581.5150.4330.5940.5970.5950.6234.1110.8561.2450.866
GeoT0.27520.9864.0902.6462.3817.0280.6630.4500.4350.3551.3903.7301.3561.0730.920
DoM1.60450.8331.6030.9191.0430.4410.2860.4390.4500.4400.8413.8120.8390.6430.677
XMGN2.30044.68623.11424.69624.8056.9410.4700.3880.4200.4169.9954.5822.9933.2203.211
GINO0.98818.74711.38811.36411.3583.1440.4370.5310.5300.5311.7622.8632.4592.4532.456
10TSol0.25610.0702.6090.6652.9941.5150.4080.4450.4680.4460.6232.0261.0770.5581.156
GeoT0.27519.0812.4410.3380.3347.0280.4050.3580.3600.3551.3902.7800.9340.3490.345
DoM1.60427.2715.0850.5450.7020.4410.4990.4140.4360.4250.8413.6881.4500.4870.546
XMGN2.30034.41023.83015.80112.7806.9410.7020.6710.6610.6599.9954.9164.0003.2322.902
GINO0.98813.32817.71312.02514.2603.1440.5260.5140.4740.4971.7622.6483.0162.3872.662
40TSol0.25619.1171.7580.4631.9171.5150.2320.2870.2980.2880.6232.1060.7100.3720.744
GeoT0.27511.7423.8020.3050.3007.0280.2640.2410.2380.2361.3901.7610.9570.2690.266
DoM1.60419.7376.7950.5380.6080.4410.3190.3340.3430.3300.8412.5111.5060.4290.448
XMGN2.30023.34222.4399.4671.7666.9410.5660.5830.5800.5649.9953.6363.6172.3430.998
GINO0.9888.67619.0111.1081.8293.1440.3030.4180.4140.4101.7621.6222.8180.6780.866
79TSol0.25615.1205.3930.3192.0701.5150.2640.2990.3090.3220.6231.9981.2690.3140.816
GeoT0.2759.9284.4530.3450.3197.0280.2650.2530.2630.2621.3901.0191.0620.3010.289
DoM1.60411.4657.5370.4970.5860.4410.3110.3170.3390.3330.8411.8881.5470.4110.442
XMGN2.30028.99223.3204.8241.7646.9410.6230.6260.5820.5879.9954.2503.8191.6751.017
GINO0.9886.86818.1750.8151.4723.1440.3210.4260.4540.4291.7621.4842.7820.6080.794
AhmedML1TSol0.25624.3894.3320.2834.04313.3390.8522.1772.0842.2961.8474.5573.0710.7683.047
GeoT0.27510.8442.9130.4540.5646.3760.8172.4902.4872.4761.3242.9762.6931.0631.182
DoM1.60424.6001.1640.5880.59311.1600.8591.1201.1011.1124.2314.5961.1420.8050.812
XMGN2.30024.6466.2086.4416.2507.3571.5701.3381.3601.3744.1132.5822.8822.9592.931
GINO0.9882.38812.18112.25911.9471.6400.8931.2911.2931.2901.2731.4603.9653.9823.925
10TSol0.25622.2416.3960.4143.17713.3390.3230.4910.7120.5211.8473.6911.7720.5431.287
GeoT0.2752.7320.5670.2940.2846.3760.4150.4390.5420.6681.3241.0650.4990.3990.436
DoM1.6046.3530.5290.6590.64311.1600.5122.5792.1102.2534.2311.8041.1681.1791.204
XMGN2.30030.4948.8498.0388.7627.3571.4961.6851.5071.5864.1136.7543.8623.4803.728
GINO0.98813.6249.7950.8131.0391.6401.7971.1421.1571.0981.2734.9483.3440.9701.068
40TSol0.25648.91513.8490.2232.98313.3390.2240.3390.3080.3291.8473.3112.1660.2620.991
GeoT0.27524.8313.0900.2410.2126.3760.2170.3030.2800.2611.3242.3240.9670.2600.235
DoM1.60428.3460.7920.5040.56211.1601.0922.1802.8123.0484.2315.5631.3141.1901.309
XMGN2.30037.51323.5934.8847.4197.3571.3261.2151.2261.2364.1137.0535.3542.4473.028
GINO0.98815.7107.0860.7590.7061.6401.7162.0211.9701.9881.2735.1923.7841.2231.184
80TSol0.25636.57915.2780.3723.29313.3390.3620.3960.3990.3621.8473.6382.4600.3851.092
GeoT0.27520.98813.5440.1970.2146.3760.3570.3040.4410.3911.3242.7392.0280.2950.289
DoM1.60429.3661.1210.4150.57111.1602.1811.9651.8521.8334.2318.0031.4840.8761.024
XMGN2.30015.90219.6552.7112.0707.3571.2270.8200.8290.8314.1134.4184.0151.4991.312
GINO0.98812.9192.8990.7680.9491.6401.1901.6281.5991.5911.2733.2962.1731.1081.229
WindsorML1TSol0.25611.1020.2530.2580.347123.9206.3356.5376.3685.6375.6318.3861.2861.2821.399
GeoT0.27512.4301.2771.4011.29077.2110.9494.2474.4294.3774.6063.4342.3292.4912.376
DoM1.60412.6450.9450.9980.879100.9055.16910.88011.21310.62612.7238.0853.2073.3443.056
XMGN2.3004.7188.2117.7667.322123.3553.22817.94020.02319.66316.8443.90212.13712.47011.998
GINO0.9885.8566.2344.0475.19216.45812.15718.67219.07319.2134.0328.43710.7898.7859.987
TTβ\betaSurrogate\multicolumn{5}{c}{Source error eSe_S}\multicolumn{5}{c}{Target error eTe_T}\multicolumn{5}{c}{Geometric mean eSeT\sqrt{e_S e_T}}
ZS ↓SC ↓FT ↓GR ↓sGR ↓ZS ↓SC ↓FT ↓GR ↓sGR ↓ZS ↓SC ↓FT ↓GR ↓sGR ↓
10TSol0.25669.1513.9250.7195.460123.92017.0426.6568.3079.4905.63134.3295.1112.4437.198
GeoT0.27510.3162.4960.3030.29677.2113.7631.2602.7703.2394.6066.2311.7730.9160.979
DoM1.60438.3171.8330.6570.789100.9053.1865.1997.2535.23412.72311.0493.0872.1832.032
XMGN2.30016.7978.8905.1074.917123.3557.6323.7014.2883.22216.84411.3225.7364.6803.980
GINO0.9884.83614.6740.7730.81216.4585.1841.5214.5555.9054.0325.0074.7251.8772.190
40TSol0.2568.94925.8310.4493.481123.9209.3849.7879.9419.9935.6319.16415.9002.1125.898
GeoT0.2753.6906.3640.3090.28877.2115.51710.2139.3627.8894.6064.5128.0621.7001.507
DoM1.60430.0921.2720.4800.616100.9053.5131.3002.4561.48012.72310.2821.2861.0860.955
XMGN2.30016.84215.7772.1712.418123.35515.43312.3325.3656.43116.84416.12213.9483.4133.944
GINO0.98810.8381.8980.8141.17916.4585.1403.5006.2324.8354.0327.4642.5782.2522.387
80TSol0.2563.02824.3980.4403.508123.92010.5049.8026.9277.6555.6315.64015.4641.7465.182
GeoT0.2752.42517.3040.2930.29777.21110.14210.2788.7969.3214.6064.95913.3361.6061.663
DoM1.60450.1481.8630.5290.475100.9058.7939.3539.0927.03412.72320.9984.1742.1931.827
XMGN2.30040.47812.3341.7951.929123.35511.93310.16513.04011.25816.84421.97811.1974.8394.660
GINO0.98810.8836.4000.7631.04716.45812.5832.0326.2074.0774.03211.7023.6062.1762.066

Table 10. Lift errors of GeoReplay across target datasets TT, target training budgets β\beta and surrogates, laid out as Table 2. The blocks give the source error eSe_S of MTSM_T^S on the 48 designs of the official DrivAerML validation split, the target error eTe_T of MTTM_T^T 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\beta. In each block, bold marks the lowest value in each row.

Notes. Errors are dimensionless relative L2L_2; lower is better. Source entries require all 48 official DrivAerML validation geometries and the same validation-selected checkpoint as the target entries. Tied values at displayed precision share the bold.

Sparse GeoReplay attempts replay every fourth global training step, with a hard alignment threshold of 0.02 and replay weight 0.5. Transolver restricts source replay gradients to its final two blocks and uses 25,000 source points; the other surrogates use all layers and their native point sampling. Target gradients update all layers. These variants differ in more than replay frequency.

Complete SAV Results

Surrogate Added Value. Surrogate Added Value (SAV) compares the adapted model with the same architecture trained from scratch on the same target simulations. For held-out simulation ii, let eadapt,ie_{\mathrm{adapt},i} and eSC,ie_{\mathrm{SC},i} denote the errors of the adapted and from-scratch models, and let eˉadapt\bar e_{\mathrm{adapt}} and eˉSC\bar e_{\mathrm{SC}} denote their means over the nn held-out simulations. The adapted model starts from the pretrained surrogate. It is the fine-tuned model (FT) in Table 11. We define

SAV=eˉSC−eˉadapteˉSC=1−eˉadapteˉSC.(6)\mathrm{SAV} = \frac{\bar e_{\mathrm{SC}}-\bar e_{\mathrm{adapt}}}{\bar e_{\mathrm{SC}}}=1-\frac{\bar e_{\mathrm{adapt}}}{\bar e_{\mathrm{SC}}}. \tag*{(6)}

For eˉSC>0\bar e_{\mathrm{SC}}>0, SAV ∈(−∞,1]\in(-\infty,1], or (−∞,100%](-\infty,100\%] when reported as a percentage. A value of 1 represents zero adapted error, zero represents equal errors, and a negative value means that the adapted model has higher error. A value below −100%-100\% means that the adapted error exceeds twice the from-scratch error. SAV is conditional on the source and target datasets, surrogate, predicted quantity, target-data budget, and training protocol. SAV is close to the accuracy and efficiency gains of Poseidon [21], which compare a pretrained PDE model with a model trained from scratch at a matched number of training samples. Poseidon’s from-scratch baseline is a different architecture (FNO), whereas SAV holds the architecture fixed and therefore isolates the value of pretraining. We measure it after training, so SAV does not by itself predict transfer before adaptation. We form a 95% interval by resampling the paired held-out simulations and recomputing Equation 6. This interval measures variation across the held-out simulations conditional on the fixed split and trained models. To prevent a few high-error simulations from determining the direction of a claim, we call transfer beneficial or harmful only when the interval excludes zero and the SAV sign agrees with n−1∑ilog⁡(eSC,i/eadapt,i)n^{-1}\sum_i \log(e_{\mathrm{SC},i}/e_{\mathrm{adapt},i}).

As a reference, we report the best per-simulation mean-field predictor in the Mean column of Table 11, which we obtain by predicting each held-out geometry’s own spatial-mean field. No mean-field predictor achieves lower error, so we read a model above this value as capturing no spatial structure beyond the field mean.

AB-UPT. AB-UPT [2] needs CFD volume meshes, so Table 11 includes it only for AhmedML. On AhmedML we also trained it with β=1\beta=1, 44, 2020, 4040 and 150150 target simulations, on the splits of the budget sweep (Table 12). Fine-tuning beats training from scratch with β=1\beta=1 and β=4\beta=4 (SAV +26.0%+26.0\% and +7.8%+7.8\% in pressure L2L_2 error), loses with β=10\beta=10, 2020 and 4040 (−37.4%-37.4\%, −98.4%-98.4\% and −45.1%-45.1\%), and matches it with β=150\beta=150 (−1.1%-1.1\%). We did not run GeoReplay with AB-UPT.

†AB-UPT requires CFD volume meshes, and we report it for AhmedML only.

β\betaFTSCSAV
10.6530.882+26.0
40.2990.325+7.8
100.1730.126−37.4
200.1080.054−98.4
400.0840.058−45.1
1500.0670.066−1.1

Table 12. AB-UPT on AhmedML: mean pressure L2L_2 test error of fine-tuning (FT) and training from scratch (SC) for each target training budget β\beta, and SAV =1−FT/SC= 1 - \mathrm{FT}/\mathrm{SC}. There is no β=80\beta= 80 run.

Proofs

Proof of Proposition 1. For the fixed parameters θ\theta, write RT(θ)=EG∼PG[mT(G)]R_T(\theta) = \mathbb{E}_{G\sim P^G}[m_T(G)] and RS,w(θ)=EG∼QwG[mS(G)]R_{S,w}(\theta) = \mathbb{E}_{G\sim Q^G_w}[m_S(G)]. Add and subtract EG∼QwG[mT(G)]\mathbb{E}_{G\sim Q^G_w}[m_T(G)] and apply the triangle inequality:

∣RT−RS,w∣≤∣EPTGmT−EQwGmT∣+∣EQwG(mT−mS)∣.(7)\lvert R_T - R_{S,w}\rvert\leq\left\lvert\mathbb{E}_{P^G_T}m_T - \mathbb{E}_{Q^G_w}m_T \right\rvert+ \left\lvert\mathbb{E}_{Q^G_w}(m_T-m_S) \right\rvert. \tag*{(7)}

For any coupling π∈Π(PTG,QwG)\pi\in\Pi(P^G_T,Q^G_w), the first term is at most E(G,G′)∼π∣mT(G)−mT(G′)∣≤LGEπdC(G,G′)\mathbb{E}_{(G,G')\sim\pi}\lvert m_T(G)-m_T(G')\rvert\leq L_G\mathbb{E}_{\pi}d_C(G,G'). Taking the infimum over couplings gives LGTdC(PTG,QwG)L_G T_{d_C}(P^G_T,Q^G_w). The conditional-shift assumption bounds the second term by δphys\delta_{\mathrm{phys}}. Adding the two bounds proves (4). □\square

The first assumption requires nearby normalized geometries to have nearby conditional target risks for the fixed surrogate. The conditional-shift assumption does not require source and target CFD labels to agree. Instead, δphys\delta_{\mathrm{phys}} explicitly permits disagreement caused by factors that dCd_C omits. A small transport term is therefore insufficient when the source and target simulation processes differ substantially.

Proof of Proposition 2. The LL-smoothness of FTF_T gives the descent lemma

FT(θ−ηd)≤FT(θ)−η⟨gT,d⟩+Lη22∥d∥2.(8)F_T(\theta-\eta d) \le F_T(\theta)-\eta\langle g_T,d\rangle+\frac{L\eta^2}{2}\lVert d\rVert^2. \tag*{(8)}

The alignment condition and λ≥0\lambda\ge0 imply

⟨gT,d⟩=∥gT∥2+λ⟨gT,gR⟩≥∥gT∥2≥0.(9)\langle g_T,d\rangle=\lVert g_T\rVert^2+\lambda\langle g_T,g_R\rangle\ge\lVert g_T\rVert^2\ge0. \tag*{(9)}

When gT≠0g_T\ne0, substituting the step-size condition in Equation (5) into the descent lemma makes the sum of its last two terms nonpositive. When gT=0g_T=0, Equation (3) gives gR=0g_R=0, so the update leaves θ\theta unchanged. Thus FT(θ+)≤FT(θ)F_T(\theta^+)\le F_T(\theta) in both cases.

The proposition uses the exact target empirical gradient. With stochastic mini-batches, the same calculation applies conditionally to the sampled target loss, while a guarantee for the full empirical risk additionally requires assumptions on gradient noise. We therefore interpret the gate as a local safeguard and make no guarantee of improved generalization.

Implementation Details

Source Model Training

From DrivAerML we use the 436 simulations of the official training split, which pretrained the released models, and the 48 of the official validation split for source evaluation. We use all 500 DrivAerNet++ simulations and the 99 SHIFT-SUV sample simulations. We use 499 of the 500 AhmedML simulations and exclude run 500, whose released force file is unavailable, so its drag cannot be checked against the integrated surface fields. For WindsorML we use the official training, validation, and test splits of the release. It lists 355 designs, but runs 350 to 354 have no published fields, which leaves 280 training, 35 validation, and 35 test simulations, and each training budget is a nested subset of the official training split.

We use the model implementations and released DrivAerML checkpoints from PhysicsNeMo-CFD version 0.0.3a0 with PhysicsNeMo 2.1.1 [38]. Each matched pair uses the same dataset adapter, surface sampling, architecture, loss, optimizer family, split, epoch budget, and checkpoint-selection rule, and differs only in initialization and in an initialization-appropriate learning rate. The fine-tuned arm begins from the released source parameters; the scratch arm begins from random initialization. We train each arm to convergence with seed 0. Each run stops after 30 consecutive epochs without a validation improvement or at an epoch cap of 300, raised to 500 for arms whose validation error was still falling at 300 epochs. We retain each surrogate’s native learning-rate schedule: Transolver and GeoTransolver use step decay every 100 epochs, DoMINO uses its released multistep schedule, and X-MeshGraphNet uses cosine decay. We reached the 300-epoch cap in stages by resuming the saved model, optimizer, scheduler, and early-stopping state; both arms of every pair follow the identical schedule. The validation score is the sum of mean pressure and wall shear stress relative L2L_2 errors. We evaluate the test simulations once, using the checkpoint with the lowest validation score over the whole run.

The fine-tuned arm uses learning rate 5×10−55\times10^{-5}, suited to continuing from pretrained parameters, and the scratch arm uses 1×10−31\times10^{-3}, because a randomly initialized model needs a larger rate and reusing the small fine-tuning rate would understate scratch performance and inflate SAV. Transolver trains in bfloat16 precision on sampled surface points; its released 200,000-point setting exceeds the memory of one 140 GB H200 during a backward pass. It samples 100,000 points per simulation for the first 100 epochs on H200 GPUs and 64,000 points afterwards on 80 GB H100 GPUs; both arms of each pair use the same sampling, and evaluation always covers the complete surface. The other architectures follow their native PhysicsNeMo surface representations and sampling conventions. GINO follows the published Ahmed-body configuration of Li et al. (2023): a 64364^3 latent grid on the normalized domain [−1,1]3[-1,1]^3, a GNO radius of 0.055, 64 FNO channels, a signed distance field of the geometry as latent input, and 100,000 sampled surface points. We pretrain it on the 436 DrivAerML training simulations for 100 epochs with Adam and a step schedule that halves the learning rate of 2.5×10−42.5\times10^{-4} at epoch 50, and it predicts pressure and wall shear stress at every surface cell. AB-UPT uses the released DrivAerML checkpoint and the released sampling configuration of 65,536 geometry points, 16,384 supernodes, and 16,384 surface and volume anchors. Its volume anchors are a uniform subsample of the cell centers of each geometry’s own CFD volume mesh; with the released DrivAerML cloud, this subsample reproduces the released zero-shot surface error on a held-out design. Both arms of AB-UPT follow its released Lion recipe, including an exponential moving average of the weights over a 25-epoch window. The complete campaign contains 40 training arms: fine-tuned and scratch runs for five architectures and four datasets. We separately run 20 zero-shot evaluations on the exact test simulations.

Before training or inference, we reconcile each dataset with the source coordinate, scale, and field-unit conventions. WindsorML uses a vertical yy and lateral zz axis and releases dimensionless pressure and friction coefficients on mesh points; AhmedML releases kinematic fields computed at a unit freestream speed; SHIFT-SUV contains quarter-scale vehicles. We convert point fields to cell fields, redimensionalize WindsorML coefficients and AhmedML kinematic fields at the dynamic pressure of the released PhysicsNeMo-CFD model wrappers (ρ=1.205 kg/m3\rho= 1.205\,\mathrm{kg}/\mathrm{m}^{3}, U=30 m/sU = 30\,\mathrm{m}/\mathrm{s}), scale SHIFT-SUV geometry by four and place it on the ground plane, and align normal and wall shear stress sign conventions with the source. We keep the WindsorML axes and body sizes as released, so these frame and scale differences remain part of the shift each surrogate faces. DrivAerNet++ requires no coordinate-scale or field-unit conversion. Force integration after conversion reproduces the available dataset force coefficients; detailed validation is retained in the experiment archive. Every method, including training from scratch, uses the released checkpoint’s PhysicsNeMo input preprocessing and output normalization statistics at both training and inference, so methods differ only in initialization and training and never in normalization.

Assets and computing infrastructure. DrivAerML {(citep)}(Ashton et al., 2024c), AhmedML {(citep)}(Ashton et al., 2024b), and WindsorML {(citep)}(Ashton et al., 2024a) are released under CC BY-SA 4.0, and DrivAerNet++ {(citep)}(Elrefaie et al., 2024) and the SHIFT-SUV sample {(citep)}(Luminary Cloud, 2025) under CC BY-NC 4.0. The DrivAerML checkpoints of Transolver, GeoTransolver, DoMINO, and X-MeshGraphNet are released under the NVIDIA Open Model License, PhysicsNeMo and PhysicsNeMo-CFD {(citep)}(Tangsali et al., 2025) under Apache 2.0, the GINO implementation {(citep)}(Li et al., 2023) under the MIT License, and AB-UPT {(citep)}(Alkin et al., 2025a) under the Emmi AI Non-Production License. We use all assets for non-commercial research. We train and evaluate on an internal cluster with NVIDIA H200 and RTX A6000 GPUs and on a cloud cluster with NVIDIA H100 80 GB GPUs.

Sparse GeoReplay. We also evaluate a sparse variant of GeoReplay that samples one source simulation from the same geometry-weighted distribution every fourth target training step and uses 25,000 source surface points. We restrict the source gradient to the final two Transolver blocks, including the output projection, and accept it only when its cosine similarity with the target gradient over those parameters exceeds 0.02. For an accepted source gradient, we match its norm to the target gradient norm over the same parameters and apply a weight of 0.5. Target gradients update all trainable parameters at every step, including steps when we skip or reject source replay.

Additional Qualitative Results

In Figure 5, we show the CFD pressure and the absolute pressure error of each method on one held-out geometry per comparison. On WindsorML, fine-tuning DoMINO removes most of the error over the rear half of the roof that remains after training from scratch, and GeoReplay has the lowest error of the three trained models. On SHIFT-SUV, fine-tuning Transolver leaves more error than training from scratch along the leading edge of the roof, and GeoReplay does not close this gap. A large reduction of zero-shot error can therefore coexist with negative transfer value.

Mixed-Dataset Training

The main experiments adapt surrogates pretrained on DrivAerML alone. As a first step toward pretraining across datasets, we train Transolver (Wu et al., [40]) and GeoTransolver (Adams et al., [1]) from random initialization on the union of DrivAerML, DrivAerNet++, SHIFT-SUV, AhmedML, and WindsorML, and compare each mixed model with a reference of the same architecture trained on a single dataset. The mixture draws each dataset as often as its reference does, so both models see the designs of a dataset equally often. Both are evaluated on the test designs of that dataset without further training (Table 13). These models use splits and a training recipe different from those of the main experiments, so the errors are not directly comparable with the main tables.

DatasetModelRefMixΔ\Delta
DrivAerMLTransolver0.0550.048−12%
GeoTransolver0.0460.040−13%
DrivAerNet++Transolver0.1750.172−1%
GeoTransolver0.1690.166−2%
SHIFT-SUVTransolver0.0500.051+1%
GeoTransolver0.0470.048+2%
AhmedMLTransolver0.0350.035+1%
GeoTransolver0.0310.033+6%
WindsorMLTransolver0.0830.082−1%
GeoTransolver0.0800.083+5%

Table 13. In-domain evaluation at equal exposure. Pressure L2L_2 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\Delta= \mathrm{Mix}/\mathrm{Ref} - 1; negative Δ\Delta 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%).

Mixed training lowers the pressure error on DrivAerML by 12% for Transolver and 13% for GeoTransolver, and by 1% and 2% on DrivAerNet++. On SHIFT-SUV, AhmedML, and WindsorML, the mixed models stay within 6% of their references in either direction, a difference comparable with the spread of up to 4% between the two Transolver seeds. Training on all five datasets therefore helps on the two DrivAer-derived datasets and matches single-dataset training on the others. This benefit requires the evaluated dataset to be part of the mixture. In an earlier leave-one-dataset-out experiment, models evaluated on a dataset excluded from their training had pressure errors 6 to 20 times those of the reference trained on that dataset, so mixed training alone did not transfer to an unseen vehicle dataset.

References

  1. [1]Corey Adams, Rishikesh Ranade, Ram Cherukuri, and Sanjay Choudhry. GeoTransolver: Learning physics on irregular domains using multi-scale geometry aware physics attention transformer. arXiv preprint arXiv:2512.20399, 2025.arxiv.org/abs/2512.20399
  2. [2]Benedikt Alkin, Maurits Bleeker, Richard Kurle, Tobias Kronlachner, Reinhard Sonnleitner, Matthias Dorfer, and Johannes Brandstetter. AB-UPT: Scaling neural CFD surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers. Transactions on Machine Learning Research, 2025a.arxiv.org/abs/2502.09692
  3. [3]Benedikt Alkin, Richard Kurle, Louis Serrano, Dennis Just, and Johannes Brandstetter. AB-UPT for automotive and aerospace applications. arXiv preprint arXiv:2510.15808, 2025b.arxiv.org/abs/2510.15808
  4. [5]Neil Ashton, Jordan B Angel, Aditya S Ghate, Gaetan K. W. Kenway, Man Long Wong, Cetin Kiris, Astrid Walle, Danielle C. Maddix, and Gary Page. WindsorML: High-fidelity computational fluid dynamics dataset for automotive aerodynamics. arXiv preprint arXiv:2407.19320, 2024a.arxiv.org/abs/2407.19320
  5. [6]Neil Ashton, Danielle C. Maddix, Samuel Gundry, and Parisa M Shabestari. AhmedML: High-fidelity computational fluid dynamics dataset for incompressible, low-speed bluff body aerodynamics. arXiv preprint arXiv:2407.20801, 2024b.arxiv.org/abs/2407.20801
  6. [7]Neil Ashton, Charles Mockett, Marian Fuchs, Louis Fliessbach, Hendrik Hetmann, Thilo Knacke, Norbert Schonwald, Vangelis Skaperdas, Grigoris Fotiadis, Astrid Walle, et al. DrivAerML: High-fidelity computational fluid dynamics dataset for road-car external aerodynamics. arXiv preprint arXiv:2408.11969, 2024c.arxiv.org/abs/2408.11969
  7. [8]Jules Berman and Benjamin Peherstorfer. CoLoRA: Continuous low-rank adaptation for reduced implicit neural modeling of parameterized partial differential equations. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 3565–3583. PMLR, 21–27 Jul 2024.arxiv.org/abs/2402.14646
  8. [9]Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021.arxiv.org/abs/2108.07258
  9. [10]Shuhao Cao, Francesco Brarda, Ruipeng Li, and Yuanzhe Xi. Spectral-Refiner: Accurate fine-tuning of spatiotemporal Fourier neural operator for turbulent flows. In International Conference on Learning Representations, pages 60348–60384, 2025.arxiv.org/abs/2405.17211
  10. [11]Arslan Chaudhry, Marc’Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with A-GEM. In International Conference on Learning Representations, 2019.arxiv.org/abs/1812.00420
  11. [12]Chris Choy, Alexey Kamenеv, Jean Kossaifi, Max Rietmann, Jan Kautz, and Kamyar Azizzadenesheli. Factorized implicit global convolution for automotive computational fluid dynamics prediction. arXiv preprint arXiv:2502.04317, 2025.arxiv.org/abs/2502.04317
  12. [13]Timothy Costa. NVIDIA AI physics transforms aerospace and automotive design, accelerating engineering by 500x. NVIDIA Blog, October 2025.
  13. [14]Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation for domain adaptation. In Ad- vances in Neural Information Processing Systems, volume 30, pages 3730–3739, 2017.
  14. [15]Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems, volume 26, 2013.
  15. [16]Yunshu Du, Wojciech M. Czarnecki, Siddhant M. Jayakumar, Mehrdad Farajtabar, Razvan Pascanu, and Balaji Lakshminarayanan. Adapting auxiliary losses using gradient similarity. arXiv preprint arXiv:1812.02224, 2018.arxiv.org/abs/1812.02224
  16. [17]Mohamed Elrefaie, Florin Morar, Angela Dai, and Faez Ahmed. DrivAerNet++: A large-scale multi-modal car dataset with computational fluid dynamics simulations and deep learning benchmarks. Advances in Neural Information Processing Systems, 37:439–536, 2024.
  17. [18]Mohamed Elrefaie, Dule Shu, Matt Klenk, and Faez Ahmed. CarBench: A comprehensive benchmark for neural surrogates on high-fidelity 3D car aerodynamics. arXiv preprint arXiv:2512.07847, 2025.arxiv.org/abs/2512.07847
  18. [19]Jan Hagnberger and Mathias Niepert. SMART: Scalable mesh-free aerodynamic simulations from raw geometries using a transformer-based surrogate model. arXiv preprint arXiv:2601.18707, 2026.arxiv.org/abs/2601.18707
  19. [20]Steve Hanneke and Samory Kpotufe. On the value of target data in transfer learning. Advances in Neural Information Processing Systems, 32, 2019.arxiv.org/abs/2002.04747
  20. [21]Maximilian Herde, Bogdan Raonić, Tobias Rohner, Roger Käppeli, Roberto Molinaro, Emmanuel de Bézenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for PDEs. In Advances in Neural Information Processing Systems, volume 37, 2024.arxiv.org/abs/2405.19101
  21. [22]James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017.DOI
  22. [23]Xuhong Li, Yves Grandvalet, and Franck Davoine. Explicit inductive bias for transfer learning with convolutional networks. In International Conference on Machine Learning, pages 2825–2834, 2018.arxiv.org/abs/1802.01483
  23. [24]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2021.
  24. [25]David Lopez-Paz and Marc’Aurelio Ranzato. Gradient episodic memory for continual learning. In Advances in Neural Information Processing Systems, volume 30, 2017.
  25. [26]Luminary Cloud. SHIFT-SUV Sample: High-fidelity computational fluid dynamics dataset for SUV external aerodynamics. Hugging Face dataset luminary-shift/SUV-sample, 2025.
  26. [27]Michael McCabe, Bruno Régaldo-Saint Blancard, Liam Parker, Ruben Ohana, Miles Cranmer, Alberto Bietti, Michael Eickenberg, Siavash Golkar, Geraud Krawczyk, Francois Lanusse, Mariel Pettee, Tiberiu Tesileanu, Kyunghyun Cho, and Shirley Ho. Multiple physics pretraining for spatiotemporal surrogate models. In Advances in Neural Information Processing Systems, volume 37, 2024.DOI
  27. [28]S Chandra Mouli, Danielle C. Maddix, Shima Alizadeh, Gaurav Gupta, Andrew Stuart, Michael W Mahoney, and Yuyang Wang. Using uncertainty quantification to characterize and improve out-of-domain learning for PDEs. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, 2024.arxiv.org/abs/2403.10642
  28. [29]Mohammad Amin Nabian, Chang Liu, Rishikesh Ranade, and Sanjay Choudhry. X-MeshGraphNet: Scalable multi-scale graph neural networks for physics simulation. arXiv preprint arXiv:2411.17164, 2024.arxiv.org/abs/2411.17164
  29. [30]Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leap: A new measure to evaluate transferability of learned representations. In International conference on machine learning, pages 7294–7305. PMLR, 2020.arxiv.org/abs/2002.12462
  30. [31]Cuong N. Nguyen, Phong Tran, Lam Si Tung Ho, Vu Dinh, Anh T. Tran, Tal Hassner, and Cuong V. Nguyen. Simple transferability estimation for regression tasks. In Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence, volume 216 of Proceedings of Machine Learning Research, pages 1510–1521. PMLR, 2023.arxiv.org/abs/2312.00656
  31. [33]Rishikesh Ranade, Mohammad Amin Nabian, Kaustubh Tangsali, Alexey Kamenev, Oliver Hennigh, Ram Cherukuri, and Sanjay Choudhry. DoMINO: A decomposable multi-scale iterative neural operator for modeling large scale engineering simulations. arXiv preprint arXiv:2501.13350, 2025.
  32. [34]Ievgen Redko, Amaury Habrard, and Marc Sebban. Theoretical analysis of domain adaptation with optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases (ECML PKDD), pages 737–753, 2017.arxiv.org/abs/1610.04420
  33. [35]David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. In Advances in Neural Information Processing Systems, volume 32, 2019.
  34. [36]Jian Shen, Yanru Qu, Weinan Zhang, and Yong Yu. Wasserstein distance guided representation learning for domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, pages 4058–4065, 2018.arxiv.org/abs/1707.01217
  35. [37]Shashank Subramanian, Peter Harrington, Kurt Keutzer, Wahid Bhimji, Dmitriy Morozov, Michael W Mahoney, and Amir Gholami. Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. Advances in neural information processing systems, 36:71242–71262, 2023.arxiv.org/abs/2306.00258
  36. [38]Kaustubh Tangsali, Rishikesh Ranade, Mohammad Amin Nabian, Alexey Kamenev, Peter Sharpe, Neil Ashton, Ram Cherukuri, and Sanjay Choudhry. A benchmarking framework for AI models in automotive aerodynamics. arXiv preprint arXiv:2507.10747, 2025.arxiv.org/abs/2507.10747
  37. [39]Zirui Wang, Zihang Dai, Barnabás Póczos, and Jaime Carbonell. Characterizing and avoiding negative transfer. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11285–11294. IEEE, 2019.arxiv.org/abs/1811.09751
  38. [40]Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for PDEs on general geometries. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 53681–53705. PMLR, 21–27 Jul 2024.arxiv.org/abs/2402.02366
  39. [41]Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, volume 33, 2020.arxiv.org/abs/2001.06782
  40. [42]Yijing Zhang, Nicholas Roberts, Tanya Marwah, and Mikhail Khodak. Breakeven complexity: A new perspective on neural partial differential equation solvers. arXiv preprint arXiv:2605.15399, 2026.arxiv.org/abs/2605.15399
  41. [43]Anthony Zhou, Amir Barati Farimani, Shirley Ho, and Rudy Morel. Cross-domain pretraining for steady-state neural CFD surrogates. arXiv preprint arXiv:2610.10398, 2026.arxiv.org/abs/2610.10398

Paper details

vlnc:W_fzrek7tt
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Computer Science

generative AI tools

View statement
Looking for related papers…
Contents