acceptodds
Under review as a conference paper at ICLR 2027

CANTO: CAD-Native Transformer Operators for AI-Aided Engineering

Abstract

Modern engineering systems, from automobiles to aircraft, are defined by precise parametric computer-aided design (CAD) models. Optimizing these designs requires iterative geometric updates followed by simulation-based validation, a process that typically begins with mesh generation. In practice, this involves discretizing the updated geometry, replacing the exact representation with a sampled approximation. This meshing step is computationally expensive, brittle under geometric variation, often dependent on manual intervention, and introduces information loss before analysis begins. Existing AI surrogates promise large computational speedups to the simulation, but inherit the same representation gap by relying on meshes, point clouds, voxels, or other sampled approximations of geometry. We introduce CANTO, a neural operator transformer that maps directly from continuous CAD geometry to physical fields, without meshing at any stage of the surrogate pipeline. We formalize CAD-to-physics surrogate modeling as operator learning between sequences of parametric NURBS patches and function spaces of physical fields. CANTO introduces a NURBS-native tokenization and predicts continuous fields and engineering quantities at arbitrary query locations. We validated the performance of CANTO on several automotive and aircraft aerodynamics industry benchmarks: the AhmedML, WindsorML, DrivaerML and HiLiftAeroML. In all cases, CANTO outperforms or is on par with the state-of-the-art for all variables for both surface and volume prediction. In particular, on HiLiftAeroML, it achieves a 19.8% relative improvement on the surface pressure prediction, without first meshing or tessellating the CAD geometry. This opens the door to AI-aided engineering systems and foundation models that operate directly on CAD, enabling real-time feedback, optimization, and inverse design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operator, Transformer, CAD native, meshless, CFD

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2EpqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CANTO, a neural operator that takes CAD geometry directly in NURBS form instead of a mesh or point cloud. Trimmed CAD surfaces are first converted into untrimmed patches. Each patch is encoded into a token from its control points, weights and knot vectors. These tokens condition an AB-UPT-style transformer that predicts surface and volume fields at arbitrary query points. The paper also introduces a framework that views the surrogate as a map from boundary patch parametrizations to solution fields. On four automotive and aircraft aerodynamics datasets, CANTO improves over AB-UPT on most fields, with notable gains on HiLiftAeroML. An inverse design study on AhmedML finds lower-drag designs, which are then verified with CFD.

Critical strengths
  1. Feeding the native CAD representation into a neural operator is a fresh and well-motivated idea. The paper also shows that it scales to industrial geometries with tens of thousands of patches.

  2. The results are strong on difficult datasets, for example about 20% lower surface-pressure error than AB-UPT on HiLiftAeroML. I also liked that the inverse-design results are checked by re-running the original CFD setup, where the surrogate is within about 2% of the simulations.

Critical weaknesses
  1. The claim of avoiding meshing is overstated. Volume anchors are still sampled from the CFD volume mesh, and since the surface branch attends to the volume branch, even surface predictions depend on mesh points. The paper does not explain where these anchors come from for a new design, for instance during inverse design. Untrimming is not free either, taking around 45 minutes per DrivAerML car.

  2. It is unclear where the gains come from. AhmedML, WindsorML and HiLiftAeroML have only six to eight design parameters, so a baseline conditioned on the design parameters is needed. This is all the more relevant because, on AhmedML, a plain MLP on fixed-size NURBS features already matches the Perceiver encoder on the surface fields (Table 2). CANTO also replaces AB-UPT's sampled surface anchors with NURBS-derived surface tokens. Part of the gain may therefore come from a different surface context rather than from the geometry encoder, and the number of these tokens is not reported.

Questions for authors
  1. Where do the volume anchors come from for a new geometry without a CFD mesh, for example in the inverse-design loop?
  2. How many surface tokens does CANTO use compared with AB-UPT's anchors, and how does AB-UPT perform with a matched context size?
  3. How does a baseline conditioned on the design parameters perform?
  4. Was AB-UPT retrained with NURBS-based sampling in Fig. 2? If not, its large error there may reflect a train/test mismatch.
5 comments

R: W1, Q1

We thank the reviewer for this helpful question, which highlighted an omission in Annex E. We will revise the annex to clarify how volume anchors are obtained for new geometries without a CFD mesh within the inverse-design loop.

For the AhmedML inverse-design experiments, no new CFD mesh was generated for each candidate. We deformed 16,384 anchor points selected uniformly from run 420’s CFD volume mesh, the dataset design closest to the centre of the parameter box. We computed displacements at 3,328 corresponding NURBS surface points, then transferred that motion into the volume: P(x)=Pref+b⊙[W(S(x)−S(xref))].P(x) = P_{\text{ref}} + b \odot \left[W\left(S(x) - S(x_{\text{ref}})\right)\right].

Here, PrefP_{\text{ref}} contains the reference volume points, and S(x)−S(xref)S(x) - S(x_{\text{ref}}) gives the displacement of each surface point.

WW is a fixed 16,384 × 3,328 interpolation matrix: each row corresponds to one volume point, and each column corresponds to one surface point. Each row contains nonzero weights only for the 64 surface points closest to that volume point in the reference geometry. These weights are proportional to the inverse squared distance and normalized to sum to one, so closer surface points contribute more strongly. Multiplying the surface displacements by WW therefore assigns each volume point a weighted average of nearby surface motion.

The factor bb smoothly fades this motion from full displacement within 0.15 m of the reference body to zero at 1 m, keeping the tunnel and far field fixed. The symbol ⊙\odot denotes elementwise multiplication.

Thus, the inverse-design loop avoids per-candidate meshing by using a fixed reference point cloud. The new trimmed variant further eliminates the need for untrimming and its associated preprocessing cost. In response to the reviewer’s feedback, we are currently investigating two alternative anchor sampling strategies: one based on the SDF, and another that follows the sampling procedure used during meshing but generates anchor points independently of the mesh itself. These strategies will allow us to evaluate performance without mesh-derived volume anchors.

R: W2, Q2

The surface context contains one token per NURBS anchor patch. The table below reports training and evaluation counts for both CANTO variants and the AB-UPT models used in the comparisons. We will include these counts in the manuscript.

DatasetCaseEval anchors per geometryTraining: dropout → anchors kept
AhmedMLUntrimmed NURBS3,3280.1 → 2,995
AhmedMLTrims3,3280.1 → 2,995
WindsorMLUntrimmed NURBS~4,5000.1 → ~4,100
WindsorMLTrims~2,5000.1 → ~2250
DrivAerMLUntrimmed NURBS~44k0.2 → ~35k
DrivAerMLTrims~34k0.75 →~8.5k
HiLiftAeroMLUntrimmed NURBS~4,5000.1 →~4,000
HiLiftAeroMLTrims~1,4000.1 → ~1250

The trimmed DrivAerML model retains approximately 8.5k surface tokens during training, compared with AB-UPT’s 16,384 surface anchors, while reducing pressure, velocity, and wall-shear errors from 3.82/5.93/7.29 to 3.78/5.28/7.06. It also uses fewer surface tokens than untrimmed CANTO, which retains approximately 35k during training. These results show that the trimmed model’s gains over AB-UPT do not require a larger training surface context.

The original untrimmed DrivAerML model uses a larger training surface context. The comparisons above are not a matched-context training ablation and do not separately quantify the effects of anchor count and CAD encoding. We will make the counts and scope of these comparisons explicit.

R: Q3

Our objective is to learn a mapping from CAD geometry to solution fields, with the longer-term aim of generalizing to novel shapes as the diversity of training data increases. Conditioning on design parameters addresses prediction within a prescribed shape family, whose scope is determined by the chosen parameterization. By operating directly on NURBS representations, CANTO provides a common geometric input format across different design families.

R: Q4

AB-UPT was not retrained with the alternative sampling strategy used in Figure 2. The experiment measures the sensitivity of its trained predictions to changing the surface-anchor distribution at inference, so its degradation can include train/test sampling mismatch.

The result identifies a practical distinction: CANTO constructs its surface tokens directly from NURBS, so changing the distribution of sampled surface points does not change its surface-context construction. We will clarify the scope of this experiment in the caption.

The complementary experiment reported by the AB-UPT authors in [1], Table 5, Section 4.6, trains with CAD-sampled surface inputs and randomly sampled volume anchors. The model trained with this alternative sampling strategy remained less accurate than the CFD-mesh-sampled model. Thus, in that study, the accuracy difference persisted with consistent training and inference sampling. The authors explain that CFD-mesh sampling concentrates anchors in highly resolved regions, which typically contain more complex flow features.

Because both surface and volume sampling change, that experiment does not isolate surface sampling alone. It nevertheless demonstrates the feasibility of inference without CFD-mesh-derived anchors, complementing our experiment on inference-time surface-sampling sensitivity.

[1] Alkin, Benedikt, et al. "AB-UPT: Scaling neural CFD surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers." arXiv preprint arXiv:2502.09692 (2025).

The anchor construction is now clear. The inverse-design loop avoids remeshing each candidate by deforming anchors from a reference CFD mesh. I would scope the meshing claim accordingly and describe Figure 2 as a test of inference-time sampling sensitivity.

The trimmed DrivAerML result with fewer training tokens is encouraging. A matched-context AB-UPT comparison would help isolate the contribution of CAD encoding, and a design-parameter baseline would test its added value within the prescribed shape families. Since these comparisons remain open, I am keeping my rating at 2.

Reviewer_eP9fOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The CANTO proposed in the current work, constructs geometry and surface tokens directly from CAD NURBS control points, weights, and knot vectors, and uses a Transformer operator to predict physical fields at arbitrary query locations of the domain. The approach offers a valuable connection between engineering design representations and physical-field prediction. Results on several engineering CFD datasets, together with CFD-validated inverse design, support its effectiveness and practical relevance.

My main reservations concern the evidence for the advantages of the geometry representation, applicability beyond external-flow CFD, and the explanation of representation invariance and certain architectural details. These issues do not outweigh the methodological value.

Critical strengths

S1. Learning operators directly from CAD has clear methodological and engineering value.

Using NURBS parameters as geometry inputs reduces reliance on the discretization of geometry (e.g. point-cloud or mesh-based) and maintains a direct connection between physical-field prediction and CAD design variables. The paper implements this idea in a model that accommodates heterogeneous control-net and knot-vector sizes.

S2. The evaluation covers engineering CFD cases with substantial practical complexity.

Results on four automotive and aircraft aerodynamics datasets demonstrate competitive performance on most evaluated tasks. The sampling experiments and CANTO-PA ablation provide additional evidence for the benefit of constructing surface tokens from parametric patches.

S3. The inverse-design experiments further demonstrate practical utility.

The paper uses the differentiable mapping from CAD parameters to physical fields for constrained shape optimization and validates the resulting improvements through CFD re-simulation. This strengthens the case for using the method in engineering design iterations.

Critical weaknesses

W1. The evidence for the advantages of the geometry representation needs further qualification.

The experiments support predictive performance and reduced dependence on surface sampling, but do not fully distinguish the exactness of the original CAD representation, approximation introduced by untrimming, and information preservation in finite-dimensional tokens. The CAE supervision is generated by numerical computations on discrete meshes, whereas the main controlled comparisons use point-cloud geometry inputs, including anchors sampled from mesh nodes. These inputs do not fully exploit mesh connectivity and adjacency. The results therefore do not sufficiently distinguish the advantages of CAD-native representations from the effects of discarding mesh structure during conversion to point clouds. The paper should clarify whether existing baselines directly exploit mesh connectivity and provide a corresponding controlled comparison.

W2. Applicability to broader industrial problems remains insufficiently established.

The paper demonstrates effectiveness on several external-flow CFD settings with substantial practical complexity, including multi-component high-lift configurations. However, components that jointly define a fluid-domain boundary have different information requirements from industrial assemblies involving entity membership, heterogeneous materials, and contact or coupling interfaces. It remains unclear whether the CAD encoding retains these attributes and what limitations apply to topology changes. Clarifying these conditions and evaluating multi-material, multi-entity, or coupled problems would strengthen industrial applicability. Such extensions may further increase the industrial application potential of the proposed method.

W3. Representation invariance and certain architectural choices are insufficiently specified.

Section 4.2 requires different patch decompositions, orderings, and parameterizations of the same solid to produce the same output, but the paper does not clearly distinguish architectural guarantees from properties learned through training. The split-resolution experiment in Appendix A.3 provides preliminary evidence but does not establish all these invariances. The relationship between multiple-decomposition training in Section 4.2 and the fixed split-resolution training setting in Appendix A.3 also needs clarification. Furthermore, Section 4.4 should specify the DiT-style block structure, whether conditioning or modulation is used, and the motivation relative to standard Transformer blocks.

Questions for authors

Q1. The presented HiLift case already includes multi-component aerodynamic geometries.

Please clarify whether the input retains patch-to-entity or subdomain membership, material properties, and contact or coupling interfaces. How would the model distinguish physical problems with identical geometric boundaries but different materials or interface conditions? The current evaluation primarily covers parametric configuration changes; what limitations apply to topology changes? A multi-material or coupled example would be a valuable extension for the proposed CANTO architecture for broader industrial fields.

Q2. Please provide the structure or equations of the DiT-style blocks in the geometry branch.

Do they retain adaptive normalization, conditioning, or residual gating? If modulation is used, what is the conditioning input? Since ordinary self-attention also enables global geometric interaction, the motivation relative to a standard Transformer block should be explained and empirically demonstrated. Clear specification would suffice if this is an implementation choice; a capacity-controlled comparison would be helpful if a performance advantage is claimed.

Q3. Please distinguish prior direct encoding of NURBS parameters, the specific Perceiver tokenizer proposed here, and its integration into physical operator learning.

Table 2 changes both the parameter representation and encoder architecture, so its intended conclusion should be clarified: handling heterogeneous inputs or demonstrating an architectural advantage. Appendix C appears to describe geometrically equivalent representations; please confirm this explicitly. To isolate the encoder’s benefit, an MLP–Perceiver comparison on the same fixed-size representation with controlled capacity and training settings would be informative.

Q4. Please distinguish the exactness of the original NURBS representation, geometric fidelity after untrimming, and information preservation in the final tokens.

Table 12 reports sampled geometric errors only for Windsor; could corresponding results be provided for the other datasets requiring untrimming? The CANTO-PA ablation is useful but does not isolate geometric approximation, sampling, and architectural effects. Which results specifically support attributing the accuracy gains to avoiding geometric discretization loss?

5 comments

R: W1

Indeed, three notions should be distinguished: the exact spline definition of the supplied CAD, the approximation introduced by untrimming, and the learned compression into finite-dimensional patch tokens.

CANTO directly encodes the supplied spline coefficients rather than first sampling the geometry into a point cloud. The exactness concerns the supplied spline representation; the patch tokens are learned features. We will replace “lossless geometric tokenization” with “direct encoding of the supplied spline representation.” The trimmed variant additionally encodes the trimming curves directly, removing the untrimming approximation while retaining strong predictive accuracy.

The untrimmed version of CANTO presented in the paper does not exploit surface connectivity; however, the trimmed version of CANTO includes a masked-attention based message passing between edges and neighboring surfaces, thus taking advantage of the shape’s topology.
The directly inspected AB-UPT implementation consumes surface coordinates, pools local point neighborhoods into supernodes, and uses attention; it does not consume the original CFD mesh connectivity. Table 1 includes Graph U-Net and GINO results imported from the AB-UPT study, but these are not a new controlled comparison against an encoder preserving the supplied CFD mesh adjacency. The MeshGraphnet architecture [1], which does exploit mesh connectivity and adjacency was benchmarked on the AhmedML and DrivaerML datasets in [2], Table 4, with relative L2 errors specified in the table below:

ModelAhmedMLDrivAerML
psuτpvpsuτpv
MeshGraphNet3.723.975.014.5014.3015.0120.1321.82

These published results provide additional benchmark context against a mesh-based architecture. We will distinguish them from comparisons conducted under our own training and evaluation protocol.

R: W2, Q1

The current CANTO model was tailored to external aerodynamics, including complex multi-component high-lift configurations. In the prescribed rigid-wall flow problems studied here, the geometry, fluid properties, and flow and wall boundary conditions specify the fluid problem. Solid-material properties governing deformation or conjugate heat transfer are not additional required inputs, since these effects are not modeled. We will clarify this scope in the manuscript.

For a multi-material extension, we would introduce material properties as conditioning inputs to DiT-style transformer blocks, following the approach used to condition AB-UPT on angle of attack for SHIFT-Wing [3]. The conditioning vector would encode relevant properties, such as Young’s modulus and Poisson’s ratio for structural mechanics or thermal conductivity for heat transfer, and modulate transformer features through adaptive normalization and residual gating.

R: W3

We agree with the reviewer that the manuscript should distinguish architectural invariance from robustness learned through training. Architecturally, the model is invariant to patch ordering, provided patch-associated data are reordered consistently. However, the NURBS representation of a surface is not unique, and different patch decompositions can describe the same geometric shape. The encoder retains control-net and knot-sequence positions and is therefore not invariant by construction to arbitrary reparameterizations, knot insertion, degree elevation, or patch subdivision.

We are conducting an experiment with shape-preserving CAD augmentations to evaluate prediction consistency across equivalent representations of the same geometry. We will report the tested transformations and measured differences, and clarify the invariance claims in the manuscript.

Section 4.2 describes invariance of the target physical operator. We will distinguish this property from the invariances guaranteed or learned by CANTO. The model in Appendix A.3 was trained at split level 4, its stable performance from levels 4 to 5 supports local robustness to patch refinement. The new augmentation experiment separately evaluates training with equivalent CAD descriptions.

R: Q3

We agree that Table 2 changes both the parameter representation and the encoder architecture, and therefore does not isolate the encoder’s architectural benefit. The intended conclusion was that even a simple MLP geometry encoder can achieve strong performance on AhmedML, where all NURBS patches admit a compact representation with a common control-net size and polynomial bi-degree.

We confirm that the fixed-size and variable-size representations described in Appendix C are geometrically equivalent: expressing the planar patches as biquadratic surfaces changes the spline degree and coefficient representation without changing their geometry.. For more complex CAD models, however, enforcing a common representation across heterogeneous patches can require substantial degree elevation and knot insertion, unnecessarily increasing the representation size. The Perceiver-based encoder accommodates this heterogeneity directly.

We will clarify this intended conclusion and extend the ablation to compare the MLP and Perceiver encoders on the same fixed-size representation, with controlled model capacity and identical training settings, to isolate the effect of the encoder architecture.

R: Q4

The sampled geometric errors for DrivaerML and HiLiftAeroML are reported in the table below. AhmedML requires no untrimming because its geometry can be represented using tensor-product NURBS patches.

Dataset (case)Bbox diag. (m)Max rel. errorMedian rel. error
DrivAerML6.142.00×10−31.5×10−7
HiLiftAeroML73.05.39×10−43.4×10−16

CANTO-PA retains the NURBS geometry encoder but uses sampled surface anchors. Figure 6 shows robustness to volume-anchor sparsity, while retaining sensitivity to surface-anchor sparsity. The comparison with full CANTO further supports the benefit of constructing surface anchors directly from NURBS patches: on the original submission’s WindsorML split, pressure, velocity, and wall-shear errors decrease from 8.04/7.30/7.55 to 6.81/6.69/5.46, respectively (Table 7). These results provide evidence for the complementary roles of the CAD geometry encoder and NURBS-derived surface anchors, although they do not separately quantify the contributions of geometric approximation, sampling, and architecture.

The new trimmed variant provides further evidence for the practical benefit of direct CAD input: it retains strong predictive accuracy while processing the supplied surfaces and trimming curves without the untrimming approximation.

We will clarify these aspects in the manuscript.

[1] Pfaff, Tobias, et al. "Learning mesh-based simulation with graph networks." arXiv preprint arXiv:2010.03409 (2020).

[2] Zhou, Hang, et al. "Transolver-3: Scaling up transformer solvers to industrial-scale geometries." arXiv preprint arXiv:2602.04940 (2026).

[3] Alkin, Benedikt, et al. "AB-UPT for automotive and aerospace applications." arXiv preprint arXiv:2510.15808 (2025).

Thanks to the authors for the clarifications. The response addresses some of my concerns, though others remain partly unresolved and some additional experiments are still pending. I appreciate that these are challenging engineering problems, and I see value in learning physical fields directly from CAD representations. I remain positive about the work and will keep my weak accept rating.

Reviewer_d2BYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CANTO learns how to map from a NURBS surface to physical fields ,queriable at arbitrary locations. The CANTO framework is evaluated on AhmedML, WindsorML, DrivAerML, and HiLiftAeroML. The accuracy is on par with or better than leading baselines on most tasks.

Critical strengths
  1. The fundamental representation question that this paper asks is an important one, and they back it up with a concrete experiment. Direct NURBS tokenization and continuous field queries connect the learned surrogate to editable CAD, while evaluation across four sizable automotive and aircraft CFD datasets tests more than a toy geometry family.
  2. The inverse-design case seems to use fresh CFD re-runs instead of relying on the surrogate predictions alone. This adds an admirable level of verification and validation to the work.
Critical weaknesses
  1. As currently presented, I don't think the paper supports the broad representation-invariance claim that the authors make. The work recognizes that different patch decompositions and parameterizations can encode the same solid, but it also shows that different patch splits can lead to substantial error (e.g., Figure 5). Additional evidence is needed, such as same-solid transformations.
  2. The tests performed in this work don't clearly show how much direct CAD input helps, or how much additional work it adds. The paper should also be clearer about what it means by independence from meshes or sampling choices. Two reported preprocessing examples take eight and ten hours, leaving the total time for a new design unclear. Until the methods are compared under the same conditions and the full preparation time is counted, the results do not show whether CAD input caused the accuracy gain or made the whole process faster.
Questions for authors
  1. For the same physical shape, do the predictions and design gradients change when its CAD description is altered without changing the shape? How large are the differences? This is necessary to help support the claims about invariance and direct CAD input.
  2. For each dataset, do the CAD files come from original designs, recreated design parameters, or fits to existing geometry? How long does preparation take per design, how often does it fail, and have the converted surfaces and normals been checked against a standard CAD tool? These answers provide evidenec to support to the claim that the method is practical.
4 comments

R: W1 and Q1.

CANTO is invariant to patch ordering when associated data are reordered consistently. The encoder retains control-net and knot-sequence positions, so invariance to reparameterization, knot insertion, degree elevation, or patch subdivision is not guaranteed by construction. We will clarify this distinction in the manuscript and report a same-shape augmentation experiment measuring prediction consistency across equivalent CAD representations.

Figure 5 changes both the patch decomposition and the number of surface-context tokens. It therefore does not isolate decomposition invariance. The plateau between split levels 4 and 5 supports local robustness near the training setting. We will revise Section 4.2 and Appendix A.3 accordingly.

R: W2

We have addressed the untrimming cost by extending CANTO to directly encode the supplied NURBS surfaces and their trimming curves. This eliminates untrimming entirely, as well as its associated runtime and geometric approximation while retaining strong predictive accuracy across all four datasets.

CANTO’s physics blocks remain the same as those of AB-UPT. The architectural changes concern the CAD geometry encoder and the NURBS-derived surface anchors, including trim-curve encoding and surface-edge message passing in the new variant. The improved results therefore support the benefit of direct CAD input and its encoding. The trimmed results additionally show that these benefits can be obtained without untrimming.

This removes the step that previously took approximately 2,800 seconds per DrivAerML design and 1,000 seconds per HiLiftAeroML design. The eight-hour and ten-hour figures in Table 13 describe preprocessing the complete datasets. Table 11 includes per-design untrimming in the original comparison with CFD.

We will also clarify the scope of mesh independence. CANTO derives geometry and surface-context tokens directly from NURBS, independently of surface-mesh sampling. In the original model, volume anchors are sampled from the CFD volume mesh, as in AB-UPT. We are conducting an additional experiment using [sampling strategy] to evaluate performance without mesh-derived volume anchors.

R: Q2

Appendix C details the geometry source for each dataset. AhmedML does not provide per-case CAD files, so we reconstruct its analytic six-parameter design family as 13 untrimmed NURBS patches—nine planar and four curved—without fitting to CFD meshes or point clouds. WindsorML provides parameterized STEP geometries with 42 parent surfaces. For DrivAerML, we received STEP geometries containing approximately 8,500 parent surfaces per design. HiLiftAeroML provides aircraft geometry (STEP format) for its parametric configurations, with angle of attack represented by rotating both the geometry and the fields.

[1] Alkin, Benedikt, et al. "AB-UPT: Scaling neural CFD surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers." arXiv preprint arXiv:2502.09692 (2025).

The trimmed CAD variant does help to substantially addresses my concern about untrimming cost, and the clarification of the invariance claim is helpful too. However, I do not yet see the promised same-shape consistency results, including the effect on design gradients, or completed results without mesh-derived volume anchors.

Thank you for your response, we have now completed the same-shape consistency (in the global comment C) and volume-anchor experiments (in the global comment D) requested in your follow-up.

For degree elevation, the mean shift in predicted surface pressure falls from 88% to 0.016% with CAD augmentation, and the shift in its design gradient falls from 188% to 0.28%. Augmentation also improves robustness to changes in patch decomposition. Splitting all 13 faces, a subdivision extent excluded from training, reduces the corresponding shifts from 2.1% to 0.25% and from 27% to 3.1%.

Replacing mesh-derived volume anchors with anchors generated from CAD and fixed refinement settings changes every reported error by at most 0.07 percentage points on WindsorML and AhmedML.

Together with the trimmed variant, these results provide the additional evidence identified in your follow-up. We will clarify the distinction between architectural invariance and robustness learned through augmentation. Full results are given in Sections C and D of our global response.

We hope these same-shape prediction and design-gradient results, together with the volume-anchor experiments and our previous response, address your remaining concerns and support a recommendation for acceptance.

We have now also completed a new experiments training from scratch on AhmedML without mesh-derived volume anchors, using both sampling procedures described in Section D. Full results are given in comment E. Together with the same-shape prediction and design-gradient results in comment C, these provide the additional experiments requested in your follow-up. We hope these additions address your remaining concerns, and we are grateful for your feedback!

Comments on the paper

6 comments
AuthorsOpenReview ↗

CANTO model with taking NURBS surfaces and trims directly as input

We thank the reviewers for their careful reading and constructive comments, and for recognizing the value of connecting native CAD representations to physical-field prediction and CFD-validated design.

We have developed a trimmed-CAD variant that removes untrimming while retaining competitive accuracy. We have also clarified the scope of representation invariance and mesh independence, and added experiments addressing volume sampling, parameter conditioning, and encoder comparisons.

The surface encoder shown in Figure 4 is retained, and the same general architecture is adapted to encode the one-dimensional NURBS trimming curves. The new variant also explicitly incorporates surface–edge connectivity through masked-attention message passing between surfaces and their adjacent edges, followed by global self-attention blocks.

As shown in the table below, this variant achieves comparable or better predictive accuracy than the original model while eliminating untrimming preprocessing.

AhmedMLDrivAerMLWindsorMLHiLiftAeroML
Methodpsuωτpsuτpsuτpsuτ
PointNet8.025.4466.0410.0923.6328.1341.85------------
Graph U-Net6.464.1553.667.2916.1317.9827.84------------
GINO7.906.2371.818.1813.0340.5821.71------------
LNO12.957.5972.4911.5020.5123.2736.44------------
UPT4.252.7315.035.807.448.7412.93------------
OFormer4.123.6315.064.604.856.648.92------------
Transolver3.452.058.224.004.816.788.95------------
Transformer3.412.096.764.034.356.218.26------------
AB-UPT3.011.906.523.883.825.937.298.267.406.885.657.068.72
CANTO untrimmed2.662.044.833.533.975.147.158.127.057.004.535.878.52
CANTO with trims2.771.904.643.553.785.287.067.196.625.214.255.878.71

All WindsorML results in this table use the official split published by the dataset authors after our submission, ensuring a consistent comparison across methods.

AuthorsOpenReview ↗

A. Matched-context AB-UPT

We thank the reviewers for their careful reading and constructive comments, which have helped us improve the manuscript.

In addition to the new trimmed variant shared in our previous response, we have now completed the requested matched-context AB-UPT comparison, the design-parameter baseline, and the same-shape prediction and design-gradient experiment. We also report the volume-anchor experiments described in our earlier response.

The trimmed variant directly encodes the original NURBS surfaces together with their trimming curves, eliminating the untrimming procedure and its associated runtime and geometric approximation. CANTO’s physics blocks remain the same as those of AB-UPT.

Here, we compare CANTO with AB-UPT at matched surface-context size to verify that CANTO’s gains are not explained by a larger number of surface tokens.

Specifically, we trained AB-UPT on WindsorML with its surface context approximately matched to CANTO. Trimmed CANTO uses an average of 2,577 surface tokens per design, ranging from 2,120 to 2,730. We therefore trained AB-UPT with 2,560 surface anchors and kept every other setting of our AB-UPT baseline: 16,384 volume anchors, 65,536 geometry points, the same optimizer and 26,250-step schedule, and the official split.

The table below reports test relative L2 errors (%) and relative drag-coefficient error.

ModelSurface contextpsp_suτ\taupvp_vCd error
CANTO with trims~2,600 NURBS-derived surface tokens7.186.615.208.514.22
AB-UPT2,560 surface anchors11.106.849.318.987.95

With approximately matched surface context, AB-UPT has substantially higher surface errors than CANTO. Its drag error is nearly twice as large. We also retrained AB-UPT with 16,384 surface anchors using the same code. CANTO achieves lower pressure and drag errors than this configuration, whose corresponding errors are 8.26% and 7.53%.

These results support the benefit of CAD input and its encoding and show that CANTO’s gains do not require a larger surface-token budget.

AuthorsOpenReview ↗

B. Baseline taking design parameters as inputs

We compare CANTO with a design-parameter baseline to test whether conditioning on the parameters controlling variations of the original CAD can match the accuracy of direct CAD input. These design parameters describe the design variations without explicitly providing the resulting geometry to the model.

Within CANTO, replacing the explicit geometry input with design parameters would leave geometric information in its NURBS-derived surface anchors. We therefore implemented Parameter-DeepONet [1], which conditions directly on design parameters. We trained it on AhmedML and DrivAerML in the same setting as the other approaches.

The table below reports relative L2 errors (%).

AhmedMLDrivAerML
Methodpsuωτpsuτ
CANTO with trims2.771.904.643.553.785.287.06
Parameter-DeepOnet12.98.107.3112.531.2717.9848.34

CANTO achieves substantially lower errors across all evaluated quantities on both datasets, including the six-parameter AhmedML family. The errors of Parameter-DeepONet are lower on the simplified AhmedML geometry and substantially higher on DrivAerML. On DrivAerML, its pressure, velocity and wall-shear errors are 31.27%, 17.98% and 48.34%, compared with 3.78%, 5.28% and 7.06% for CANTO.

These results support the benefit of providing the CAD geometry directly. Since the models also differ in architecture, this comparison does not isolate the contribution of geometry conditioning alone.

[1] Shukla et al. “Deep neural operators can serve as accurate surrogates for shape optimization: a case study for airfoils.” 2023.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.