WorksComputer Science
GeoTransolver: Learning Physics on Irregular Domains Using Multi-scale Geometry Aware Physics Attention Transformer
Abstract
We present GeoTransolver, a multiscale geometry-aware physics attention transformer for Computer Aided Engineering (CAE). GeoTransolver extends the Transolver backbone with GALE (Geometry-Aware Latent Embeddings) attention, which pairs physics-aware self-attention on learned state slices with cross-attention to a shared geometry and global context computed via multi-scale ball queries (inspired by Domino) and reused in every block. Implemented and released in NVIDIA PhysicsNeMo, GeoTransolver persistently projects geometry and global parameters, into physical state spaces to anchor computations to domain structure and operating regimes. We benchmark on DrivAerML, SHIFT-SUV, and SHIFT-Wing against Domino, Transolver (PhysicsNeMo implementation), and literature-reported AB-UPT, evaluating drag/lift R2 and relative L1 errors on field variables. As an additional nonlinear structural mechanics application, we also report Transolver and GeoTransolver results on bumper-beam and full-vehicle Body-in-White (BIW) crash-dynamics benchmarks, evaluating relative L2 trajectory error and probe-level kinematic MSE. GeoTransolver delivers improved accuracy, robustness to geometry and regime shifts, and favorable data efficiency; we include DrivAerML ablations and qualitative contour and design-trend results, advancing operator learning for high-fidelity surrogates on complex, irregular, non-linear domains.
Introduction
Computer-aided engineering (CAE) increasingly depends on AI-based surrogates to accelerate design exploration and reduce reliance on costly high-fidelity simulation [1, 2, 3, 4]. Despite rapid progress in neural operators and physics-informed architectures, deploying these models on industrial CAE workloads exposes three persistent failure modes that limit their trustworthiness for engineering decision-making:
Geometry dilution across depth. Industrial meshes contain millions of surface and volume elements with heterogeneous refinement near boundaries, wakes, and separation regions. In transformer-based operators that inject geometric information only at the input layer, boundary and shape detail is progressively diluted as latent representations evolve through successive attention blocks. Later layers lose access to fine-grained domain structure, degrading predictions precisely where near-wall physics matters most – for example, wall-shear stress near the A-pillar or wheel arches of a vehicle, or contact surfaces in crash dynamics [5, 6].
Preprint.
Scale mismatch on non-uniform discretizations. Automotive and aerospace meshes exhibit orders-of-magnitude differences in cell size between boundary-layer regions and far-field elements. A single-radius neighborhood either floods with boundary-layer detail or misses it entirely; global attention with uniform tokens averages over the disparity. Capturing both near-boundary gradients and far-field coupling requires explicit multi-scale geometric aggregation that spans both regimes simultaneously.
Single-shot regime conditioning. CAE surrogates must generalize across operating regimes (Mach number, angle of attack, impact velocity) and boundary conditions. Existing conditional operator learners typically concatenate regime parameters to input tokens – a single conditioning event that does not persist as the latent state evolves through depth. Just as U-Net skip connections were introduced to prevent high-resolution spatial features from being lost through successive pooling layers [7], persistent conditioning is needed to prevent geometry and regime information from being washed out through successive attention layers. Without it, the model’s internal state becomes decoupled from the imposed operating conditions in deeper layers.
This paper introduces GeoTransolver, a transformer architecture for CAE in which each design choice directly addresses one of these domain-specific failure modes. GeoTransolver builds on the Transolver [6] backbone and introduces Geometry-Aware Latent Embeddings (GALE) attention, which pairs physics-aware self-attention on learned state slices with cross-attention to a shared geometry and global context at every block. To address scale mismatch, multi-scale ball queries with radii spanning boundary-layer to far-field scales extract local neighborhood features; geometry and global operating parameters are then projected into the same latent slice-token space used by the attention layers, providing persistent conditioning that prevents both geometry dilution and regime drift. Prior methods condition on geometry only at the input [8, 9] or at encoder/decoder boundaries [10]; GeoTransolver maintains geometry and regime context in the same learned basis as the evolving latent state at every layer.
GeoTransolver is implemented and released as open-source software in the NVIDIA PhysicsNeMo framework [11], which provides geometry data pipelines, multi-physics conditioning utilities, and scalable training/inference infrastructure necessary for high-resolution CAE workloads.
We benchmark GeoTransolver against Domino [8], AB-UPT [9], and Transolver [6] on three industrial CAE datasets: DrivAerML [12] (500 parametric vehicle variants on 140M-element meshes), Luminary SHIFT-SUV [13], and Luminary SHIFT-Wing [14], using the benchmarking framework in PhysicsNeMo [15]. These tasks span aerodynamic flows over parameterized vehicle and wing geometries across diverse operating regimes. We evaluate field-level reconstruction (velocity, pressure, wall shear) and integral quantities (drag and lift coefficients), along with generalization to unseen geometries and design trends. We further include crash-dynamics benchmarks (bumper-beam impact and full-vehicle Body-in-White frontal impact) to evaluate whether geometry-aware conditioning transfers beyond external aerodynamics to transient, contact-driven structural mechanics with large deformations, and to demonstrate that the geometric context pathway can be paired with alternative efficient attention backends [16].
Our contributions are two-fold:
A geometry-persistent transformer design (GeoTransolver) in which each architectural choice addresses a specific failure mode: GALE cross-attention at every block prevents geometry dilution (failure mode 1); multi-scale ball queries with radii spanning boundary-layer to far-field scales resolve the scale mismatch (failure mode 2); and a global-parameter context projector that injects operating-regime information at every depth prevents regime drift (failure mode 3). The geometry-aware context enters through a modular cross-attention path, so the self-attention backend can be swapped – we demonstrate this by replacing physics attention with FLARE [16] on crash-dynamics benchmarks.
Industrial-scale evaluation and open-source release within NVIDIA PhysicsNeMo, benchmarked on DrivAerML (500 vehicle variants, 140M-element meshes), SHIFT-SUV, SHIFT-Wing, and two crash-dynamics datasets. GeoTransolver delivers improved accuracy and robustness across all benchmarks, with ablations validating each design choice independently.
Background and related work
AI-based surrogates for CAE span operator learning, mesh/geometry-aware representations, multi-scale architectures, and physics-informed training. Below we survey the most relevant lines of work and position GeoTransolver relative to them.
Neural operators
Neural operators learn mappings between function spaces for resolution-agnostic PDE surrogates [17]. DeepONet [18] introduced branch–trunk architectures; FNO [19, 20] uses spectral convolution for nonlocal interactions; and GNO [21, 22] targets irregular discretizations via message passing. Transformer-based operators [6, 23, 24, 25, 26, 27, 28] improve long-range coupling via global attention over latent fields. GeoTransolver replaces standard attention with GALE, structuring self-attention around physics-aware state slices and cross-attending to geometry and global context at every block.
Geometry and mesh-aware encoders
MeshGraphNet and related GNNs [29, 30, 31, 32] popularized message passing on simulation meshes; extensions add learnable stencils and multigrid coarsening [33, 34]. Point-cloud networks [35, 36, 37] introduced multi-scale local aggregation for unstructured geometry. Domino [8] demonstrated that multi-scale ball queries paired with global shape descriptors improve generalization across shape families. AB-UPT [9] explores unified physics transformers with geometry-grounded tokens.
GAOT [10] is the most closely related prior work: it combines multiscale attentional graph neural operator (GNO) encoders and decoders with geometry embeddings, achieving strong results on DrivAerML and other industrial-scale datasets. A concurrent 3D extension [10] further scales this approach. The key architectural difference is where geometry enters the network: GAOT injects geometric information at encoder and decoder boundaries, so the transformer processor layers between them operate without direct geometric recall. GeoTransolver instead projects geometry into the same slice-token basis used by the latent state and cross-attends at every layer, providing persistent conditioning throughout the forward pass. The two approaches also differ in metrics (GAOT reports MSE/MAE; GeoTransolver reports relative and ), precluding direct numerical comparison on shared datasets at this time. Beyond geometry-aware encoders, U-Net hierarchies, graph coarsening, and multigrid networks [38, 39, 40] and spectral methods [19, 41] also target multi-scale coupling, the latter struggling on irregular domains. GeoTransolver addresses the scale-mismatch failure mode by using multi-scale ball-query neighborhoods while preserving a dedicated global-context channel via cross-attention at every depth.
Physics-aware conditioning and constraints
PINNs [42] and PINO [43] enforce PDEs via residual losses; equivariant architectures [44, 45, 46] encode symmetries (e.g., ); and conditional operator learning introduces regime metadata but typically conditions only once, risking representation drift. GeoTransolver projects geometry and global context into aligned physical-state tokens at every block for persistent conditioning, and is compatible with PINN/PINO-style losses.
AI modeling for CAE
Industrial CAE surrogates must handle irregular, heterogeneous geometries (vehicles, wings, structural assemblies), capture multiscale physics from boundary layers to far-field wakes, and remain conditioned on operating regimes across design families. As identified in § 1, these requirements converge on three failure modes that expose fundamental gaps in current architectures: geometry dilution across depth, scale mismatch on non-uniform discretizations, and single-shot regime conditioning. No single prior method addresses all three.
Table 1 maps each baseline to the three failure modes. Methods that encode geometry only at the input (Domino, AB-UPT) or at encoder/decoder boundaries (GAOT) leave later transformer layers without direct geometric recall. Methods without multi-scale neighborhoods (FNO, DeepONet, Transolver) cannot resolve the 1000:1 resolution ratios typical of industrial meshes. No prior method provides persistent context conditioning at every layer in the same latent basis as the evolving physical state. GeoTransolver is designed to close all three gaps simultaneously.
| FNO | DeepONet | Transolver | Mesh GraphNet | DoMINO | AB-UPT | GAOT | Geo Transolver | |
| Geometry encoding | spectral grid | branch net | input tokens | edge features | ball queries | geo tokens | GNO (multi-sc.) | ball queries + projector |
| Multi-scale | — | — | — | local | multi-radii | ✓ | ✓ | multi-radii |
| Context depth | — | — | — | — | input only | input only | enc/dec only | every layer (x-Attn) |
Table 1. Failure-mode coverage across methods. Only GeoTransolver provides persistent geometric context at every transformer layer.
GeoTransolver
GeoTransolver takes three inputs – query points with local field values, a geometry surface mesh, and optional global operating parameters (e.g., Mach number, angle of attack) – and predicts output fields at the query positions. Each component of the architecture addresses one of the three failure modes from § 1: a shared context encodes geometry and global parameters in a latent slice-token space and is reused at every layer to prevent geometry dilution; multi-scale ball queries within the context construction resolve boundary-layer-to-far-field scale mismatch and additionally augment input tokens with local geometric features; and a global parameter projector injects regime information into the same slice-token space, providing persistent regime conditioning across depth. Local field embeddings then pass through a stack of GALE (Geometry-Aware Latent Embedding) blocks, each performing physics-aware self-attention on learned slices followed by cross-attention to the shared context, blended via a learnable gate (Figure 1).
Figure 1. GeoTransolver architecture. Right: the context projector constructs a shared geometry and global context by projecting multi-scale ball-query features and global parameters into slice-token space via a half-attention operation; learned geometric features are also concatenated onto local input tokens. Center: The learned context is continually injected to the model to provide persistent geometrical recall. Left: each GALE block extends Transolver’s Physics Attention with the addition of cross attention to the shared global context.
Notation. Throughout this section we use: for the number of query points, for geometry points, for attention heads, for slices (latent states) per head, for head dimension, for input feature dimension, for global parameter dimension, for geometric processor output dimension, and for the number of ball-query scales.
GeoTransolver inputs
GeoTransolver operates on a set of query points, input Geometry , and global operating parameters :
Query points with positions and features:
where may include surface normals, signed distance fields, Fourier embeddings, or any task-specific encoding. The model predicts output fields at each query position .
Geometry point cloud , , representing the surface mesh or boundary of the domain.
Global parameters (e.g., Mach number, angle of attack, Reynolds number).
Both the geometry and global parameters are projected into the shared geometry context as described in § 3.3, which continuously augments the model’s latent space.
Persistent geometry recall via GALE
Transolver’s physics attention [6] avoids the quadratic cost of full attention by projecting input tokens onto a small set of learned slices (latent basis vectors) and performing self-attention in that reduced space, scaling to input sizes where full self-attention is infeasible.
This design is efficient but introduces the geometry dilution failure mode identified in § 1: because each layer receives only the output of the previous layer, geometric and boundary information present in the original inputs isz progressively diluted with depth. Later layers have no mechanism to recall fine-grained domain structure, degrading predictions in near-wall regions where boundary-layer physics depends on local geometry. GALE (Geometry-Aware Latent Embeddings) addresses this directly by adding cross-attention from the slice tokens to a shared context that encodes geometry and global parameters in the same slice-token space, ensuring every layer retains direct access to domain structure.
The key deviations for a GALE layer occur after the latent state self attention, as seen in Figure 1 (left). The learned projections for a given layer are used in a cross-attention mechanism with the ‘context’ embedding, learned directly from the input points and geometries. This cross-attention is then blended with the self-attention output via a learnable scalar gate , where is a per-layer parameter and is the sigmoid function. This allows the model to adapt its balance between geometric context and self-attended state at each depth. After mixing, the slice tokens are de-sliced back to the input dimensionality as in Transolver [6].
Although we use Transolver’s physics attention as the default self-attention backend, the GALE formulation is not restricted to this choice. The geometry-aware context enters through a separate cross-attention path, so the self-attention branch can be replaced by other efficient token-mixing mechanisms. In the crash-dynamics experiments, we include a lightweight backend-swap demonstration in which the physics-attention backend is replaced with FLARE attention [16] while keeping the GeoTransolver geometry-context pathway unchanged.
Multi-scale context for boundary-layer to far-field coupling
The second and third failure modes – scale mismatch and single-shot regime conditioning – are addressed by the context construction stage. Before any GALE block is applied, GeoTransolver builds a shared context that encodes the geometry and global parameters in the same slice-token space used by GALE attention. The context is assembled from two reusable building blocks – a context projector and a geometric processor – composed in three independent paths: a geometry tokenizer, a global tokenizer, and a multi-scale feature extractor (see Figure 1, right). Each path produces slice tokens in the same latent space as the GALE attention slices; the outputs are concatenated into a single context tensor , computed once per input and reused at every depth.
The multi-scale extractor also produces geometric embedding for each input query that are concatenated onto each input token before the first GALE block, giving every point direct access to nearby boundary geometry information.
Context projector
The context projector implements the encoding half of Transolver’s physics attention [6]: it maps a variable-length token sequence into a fixed set of slice tokens in the same latent space as the GALE attention slices, but never decodes back. Given tokens , learned projections yield . A per-head map produces slice logits ; temperature-scaled softmax and normalised aggregation with yield slice tokens :
where denotes column-normalised . The same architecture is instantiated with separate learned parameters for each context source below.
Geometric processor
A geometric processor extracts local geometric context at scale . For each point in , a ball query retrieves up to neighbours within radius from ; the flattened coordinates are mapped by a learned MLP to a bounded embedding:
Each geometric processor is parameter-light and independently parameterised; one is instantiated per scale and per cross-cloud direction.
Context construction
The three context paths below each produce slice tokens via the context projector (Eq. (2)); the results are concatenated along the feature dimension to form the shared context .
Geometry and global tokenizers. The geometry point cloud and the global operating parameters (e.g. Mach number, free-stream velocity) are each projected into slice-token space by independent context projectors:
No de-slicing is performed; both outputs remain in slice-token space for cross-attention context at every GALE layer.
Multi-scale feature extractor. The multi-scale feature extractor applies geometric processors at scales , where short-range radii capture boundary-layer detail and long-range radii enable non-local coupling (§ 5.1). Each scale uses in two symmetric cross-cloud directions and a per-scale context projector .
Input-to-geometry (context path). For each geometry point , nearby input points are gathered, processed, and tokenized into slice space:
Geometry-to-input (local augmentation path). Symmetrically, for each input point , nearby geometry points are gathered:
Per-scale features are concatenated to form a local augmentation vector , which is appended to each input token before the first GALE block, giving every query point direct access to nearby boundary geometry.
Assembled context. The shared context concatenates all slice tokens along the feature dimension:
where and the concatenation is along the last (feature) dimension. All three sources share the same slice count and head count .
Separately, the local augmentation vectors from the geometry-to-input path (Eq. (9)) are concatenated onto each input token after the initial MLP projection and before the first GALE block:
Together, provides persistent geometric and global context via cross-attention at every GALE layer, while enriches the initial per-point representation with local boundary information. Both are computed once per forward pass.
Experiment details
Datasets
The benchmarking, validation and ablation studies are carried out on 3 datasets, DrivAerML [12], Luminary SHIFT-SUV [13] and Luminary SHIFT-Wing [14]. As an additional example beyond the three aerodynamics datasets, we evaluate GeoTransolver on two crash-dynamics benchmarks: a bumper-beam impact dataset and a full-vehicle Body-in-White (BIW) frontal-impact dataset. Additional details about each of these datasets with information such as training, validation and testing splits are provided in the Appendix.
Models
We evaluate GeoTransolver on SHIFT-SUV and SHIFT-Wing against Transolver [6], DoMINO [8], and literature-reported AB-UPT [47], and probe architectural sensitivity via an ablation on DrivAerML (§ 5.1). Additional configuration and training details are provided in the appendix. For the crash-dynamics benchmarks, we compare Transolver, GeoTransolver, and GeoTransolver with a FLARE attention [16] backend under matched training protocols.
Metrics
We evaluate field-level accuracy via the relative norm , and for surface predictions report drag and lift computed from integrated surface forces . For crash dynamics we report relative error over the spatiotemporal structural response and probe-level kinematic MSE (details in Appendix).
Results
DrivAerML and ablation studies
Table 2 reports relative errors (%) on the DrivAerML test set (48 designs) and consolidates ablations over three architectural axes: GALE depth, multi-scale ball-query radii, and ball-query kernel size. Each group varies one axis while holding others at the best-found setting; bold marks the overall best row, which achieves below-5% surface and volume field error and near-perfect drag/lift agreement.
| Axis | Setting | Size | ||||||
| Layers | 6 | 9M | 3.52 | 5.88 | 0.996 | 0.987 | 3.79 | 4.44 |
| Layers | 10 | 14M | 3.25 | 5.51 | 0.995 | 0.987 | 3.31 | 4.21 |
| Layers | 14 | 20M | 3.11 | 5.29 | 0.995 | 0.983 | 3.34 | 4.24 |
| Layers | 18 | 26M | 2.95 | 5.19 | 0.994 | 0.987 | 3.08 | 4.08 |
| Layers | 20 | 29M | 2.86 | 4.9 | 0.996 | 0.991 | 3.09 | 4.02 |
| Radii | 0.05 | 12M | 3.14 | 5.38 | 0.993 | 0.989 | 3.60 | 4.34 |
| Radii | 2.5 | 12M | 3.09 | 5.38 | 0.995 | 0.986 | 3.24 | 4.20 |
| Radii | 0.05–2.5 (4) | 21M | 3.03 | 5.23 | 0.993 | 0.989 | 3.06 | 4.06 |
| Radii | 0.01–5.0 (6) | 29M | 2.86 | 4.9 | 0.996 | 0.991 | 3.09 | 4.02 |
| Kernel | 13M | 3.12 | 5.32 | 0.994 | 0.984 | 3.37 | 4.06 | |
| Kernel | 18M | 3.07 | 5.27 | 0.993 | 0.989 | 3.12 | 4.11 | |
| Kernel | 29M | 2.86 | 4.9 | 0.996 | 0.991 | 3.01 | 4.02 |
Table 2. Relative errors (%) on DrivAerML test geometries across three ablation axes. are values.
Across all three axes, accuracy improves monotonically with capacity, with diminishing returns. The best configuration (20 layers, 6 radii spanning –5.0, ; 29M params) achieves the bold-row results. Fine radii aid near-wall detail while large radii capture global couplings; larger kernels enrich local geometric aggregation. A separate study on query and geometry token counts (Appendix Table 6) shows that adequate geometric coverage is critical for volume accuracy.
Computational cost. On H100 with batch size 8, single-scale ball queries add 1.1–1.2× throughput overhead and the 4- and 6-scale multi-radii configurations add approximately 2×; the 6-scale config requires reducing query points from 200k to 160k to fit within H100’s 80 GB. Full per-configuration training and inference throughput data are in Appendix § 7.§ 4, Table 5.
| Config | Radii | Surface Train (s/step) | Surface Infer (s/step) | Volume Train (s/step) | Volume Infer (s/step) |
| No local feat. | — | ||||
| 1 scale () | [16] | ||||
| 1 scale () | [64] | ||||
| 4 scales | [16,32,32,64] | ||||
| 6 scales* | [8,16,32,32,64,128] |
Table 5. Training and inference throughput on DrivAerML (8×H100, batch size 8, bfloat16, torch.compile). Mean ± std reported.
SHIFT-SUV, SHIFT-Wing, and crash dynamics
Table 3 reports relative errors on the volume fields (pressure and velocity ) across all four SHIFT test conditions: SHIFT-SUV Estate and Fastback bodies, and SHIFT-Wing at Mach 0.5 and 0.85. GeoTransolver attains the lowest error on every column except Estate , where AB-UPT is lower by 0.0001. On surface metrics (Appendix Table 7), GeoTransolver leads on most fields with two exceptions: surface pressure at Mach 0.85 (AB-UPT) and wall-shear stress at Mach 0.5 (DoMINO, trained per-Mach). For SHIFT-SUV drag and lift , GeoTransolver achieves the highest scores (Estate =0.98/0.88, Fastback 0.99/0.99); on SHIFT-Wing all models reach =1.0 on at both Mach numbers, so integral force metrics do not discriminate.
| Model | Estate | Estate | Fastback | Fastback | Ma=0.5 | Ma=0.5 | Ma=0.85 | Ma=0.85 |
| GeoTransolver | 0.0026 | 1.36 | 0.0023 | 1.30 | 0.022 | 1.92 | 0.099 | 2.00 |
| AB-UPT | 0.0025 | 2.25 | 0.0024 | 2.21 | 0.027 | 9.56 | 0.125 | 9.51 |
| DoMINO | 0.0062 | 8.14 | 0.0067 | 7.73 | 2.25 | 21.34 | 3.17 | 29.2 |
| Transolver | 0.004 | 1.87 | 0.0039 | 1.81 | 1.19 | 3.60 | 1.34 | 3.72 |
Table 3. Volume-field relative errors (%) across SHIFT-SUV (Estate, Fastback) and SHIFT-Wing (Ma=0.5, Ma=0.85). DoMINO is trained per-Mach on SHIFT-Wing; other models share the combined dataset. Surface-field counterparts are in Appendix Table 7.
On the crash-dynamics benchmarks (Table 4), GeoTransolver reduces relative error vs. Transolver on both bumper-beam and BIW frontal-impact, and replacing the physics-attention backend with FLARE further improves both, demonstrating that the geometry-aware context pathway is backend-agnostic. Contours, design trends, deformation evolution, and probe-level kinematic MSE are reported in the Appendix.
| Model | Bumper Beam | BIW Crash |
| Transolver | ||
| GeoTransolver | ||
| GeoTransolver with FLARE |
Table 4. Crash-dynamics relative error on bumper-beam and BIW frontal-impact datasets (lower is better).
Conclusion
Industrial CAE surrogate modeling demands that transformer-based operators retain geometry awareness across depth, resolve multi-scale features on non-uniform meshes, and maintain regime conditioning throughout the forward pass. In this work, we identified these three failure modes in existing architectures and introduced GeoTransolver, a transformer design in which each component directly addresses one: GALE cross-attention prevents geometry dilution, multi-scale ball queries resolve boundary-layer-to-far-field scale mismatch, and a global-parameter context projector prevents regime drift.
We benchmarked GeoTransolver on DrivAerML, SHIFT-SUV, SHIFT-Wing, and two crash-dynamics datasets, demonstrating improved accuracy and robustness compared to Domino, Transolver, and AB-UPT. Ablations on DrivAerML validate each design choice independently, confirming that multi-scale radii, deeper GALE stacks, and adequate geometric token coverage each contribute measurably. The crash-dynamics results further show that geometry-aware conditioning transfers beyond aerodynamics to transient, contact-driven structural mechanics, and that the GALE context pathway is compatible with alternative attention backends (FLARE). GeoTransolver is released as open-source software in NVIDIA PhysicsNeMo, making it accessible to CFD researchers and application engineers.
Limitations
GeoTransolver has several limitations that frame the scope of the results presented here. The multi-scale ball-query context pathway increases per-step training and inference cost relative to baseline (Table 5, ). The DrivAerML out-of-distribution designs (lowest and highest drag) are never seen during training and remain the hardest cases (see Appendix § 7.1.1); accuracy on these extremes can likely be improved with augmented training data but is not directly addressed here. Finally, the crash-dynamics evaluation is limited to two datasets (bumper beam and a single BIW frontal-impact configuration) over short impact horizons, and broader transferability of the geometry-aware context is an active research area.
Broader impacts
GeoTransolver targets industrial CAE surrogate modelling. On the positive side, faster surrogate-based design iteration can reduce the number of high-fidelity CFD/FEM runs required during product development, lowering the HPC energy and carbon footprint of engineering workflows, and the open-source release in NVIDIA PhysicsNeMo lowers the barrier to high-fidelity CAE for academic and industrial users. On the negative side, learned surrogates that look qualitatively correct can be misused in safety-critical decisions – e.g., crashworthiness or structural sign-off – if their predictions are treated as ground truth without validation against high-fidelity simulation or experiment.
Model configuration and training details
Datasets
The benchmarking, validation and ablation studies are carried out on 3 datasets, DrivAerML [12], Luminary SHIFT-SUV [13] and Luminary SHIFT-Wing [14]. Below we provide details about each of these datasets with information such as training, validation and testing splits.
DrivAerML
DrivAerML [12] is a public, high-fidelity dataset tailored for AI-based surrogates in automotive external aerodynamics. It comprises 500 parametrically morphed variants of the DrivAer Notchback, providing broad geometric diversity for studying drag and flow behavior at scale. Each case is simulated with a scale-resolving hybrid RANS/LES approach representative of industry practice [48, 49, 50, 51], on extremely large meshes—on the order of 140 to 150 million volume elements and roughly 9 to 10 million surface points/elements [12, 52].
For every geometry, time-averaged fields are released in VTK formats: VTP files contain surface quantities (pressure and wall shear stress), and VTU files contain volume fields (velocity, pressure, and turbulence-related variables such as vorticity or turbulent viscosity). The geometry is exported as a coarse STL with around 0.3 million points. The dataset was created to address the lack of open-source, large-scale CFD data suitable for high-fidelity ML research in automotive aerodynamics.
In our study, we adopt a drag-aware split strategy. 10 % of the samples are held out for testing, and approximately 20% of this test set is designated out-of-distribution (OOD) based on drag ranges. These OOD cases correspond to some of the lowest and highest drag configurations and are not exposed during training, enabling assessment of generalization to extreme geometries and regimes. Details related to the exact training and validation split may be found in PhysicsNeMo [11] here: https://github.com/NVIDIA/physicsnemo-cfd/tree/main/workflows/bench_example/drivaer_ml_files.
Luminary SHIFT-SUV
SHIFT-SUV is an open-source, high-fidelity database of external aerodynamics developed by Luminary Cloud in collaboration with Honda, comprising thousands of transient simulations of parametrically morphed variants of the AeroSUV platform from FKFS (Forschungsinstitut für Kraftfahrwesen und Fahrzeugmotoren Stuttgart) [53]. AeroSUV shares design lineage and geometric characteristics with the widely used DrivAer platform in sedan aerodynamics [54]. Distributed under a CC-BY-NC license and hosted on HuggingFace [13], SHIFT-SUV is purpose-built for high-fidelity aerodynamic inference without requiring CFD or meshing expertise, supporting both surface- and volume-based surrogate training, real-time inference, and exploration of shape–performance correlations.
Geometry generation uses a deformation-cage approach implemented in ANSA (BETA-CAE Systems). Cage vertex motions are structured to emulate vehicle design parameters familiar to stylists and defined with guidance from Honda; the parameters are non-orthogonal (multiple parameters may affect shared control points). Configuration options—such as body style (Estate vs. Fastback) and underbody detail (smooth vs. detailed)—are treated as non-parametric toggles. Regions that should not vary across designs (e.g., wheels, tires, suspension) are excluded from morphing to preserve consistency. Specific parameter values are selected via Latin hypercube sampling. The initial dataset varies geometry while keeping boundary conditions fixed. The deformation space uses a morphing-cage workflow to apply stylist relevant surface translations relative to the baseline model, with ranges specified as target displacements in millimeters that guide, but do not strictly enforce, the final geometry.
The geometric parameters are varied as follows. Because the morphing cage couples nearby regions, these bounds serve as design targets for the cage control points rather than exact, per-point displacement limits on the surface. Additional details may be found in the dataset provided by Luminary et al. [13].
Front end changes include hood height (−50 to +50 mm), front (FR) overhang (−150 to +150 mm), windshield angle implemented via forward/backward shifts (−150 to +100 mm), and planview adjustments near the front (−75 to +75 mm),
Global and chassis-related shifts cover overall vehicle height (−150 to +150 mm) and ride height (−30 to +30 mm),
Rear end modifications include backlight angle via surface translation (−100 to +200 mm), rear (RR) overhang (−150 to +100 mm), and tapering at both the windshield and rear (−100 to +100 mm and −90 to +70 mm, respectively).
Variants are organized into four groups by scale and body style—full-scale Estate, full-scale Fastback, quarter-scale Estate, and quarter-scale Fastback—with 998 samples in each group. Quarter-scale cases validate the setup against FKFS wind-tunnel experiments for a 1/4-scale model [55], while full-scale cases follow the same methodology with refinements informed by best practices from AutoCFD [56].
Each sample has a unique volume mesh created in ANSA, yielding hex-dominant and polyhedral meshes. A grid refinement study for the validation configuration selected settings of approximately 45 million cells as a balance between accuracy and cost. All cases were run on the Luminary Cloud platform using a GPU-native, second-order (space and time) finite-volume solver [56, 57]. The turbulence model is transient delayed detached eddy simulation (DDES)—a scale-resolving DES variant—with a shear-layer–adapted length scale and a vortex tilting measure (VTM) to mitigate grey-area effects. The Spalart–Allmaras (SA) model is employed in RANS regions, and a hybrid centered/upwind convective scheme with proprietary blending limits dissipation in LES regions. An advanced shielding function prevents modeled stress depletion, avoiding premature separation. Boundary conditions include a rolling road (translating floor) and rotating wheels, with all other vehicle surfaces treated as no-slip walls. For the full-scale datasets—used in our training—the inflow is a uniform 30 m/s [56, 57].
For training we use 1996 simulations from the full-scale dataset split randomly into 80/10/10. We match the exact simulations in each split with the AB-UPT benchmarking work in Alkin et al. [47].
Luminary SHIFT-Wing
SHIFT-Wing is an open-source database centered on the NASA Common Research Model (CRM) [58] for high-speed transonic transport aerodynamics [14]. The CRM spans a wide range of configurations—from cruise to high-lift with deployed flaps/slats, optional nacelles/pylons, and empennage—and has been extensively studied experimentally and through community CFD efforts (e.g., AIAA Drag Prediction and High-Lift workshops). Developed in collaboration with Otto Aviation, the dataset is purpose-built for high-fidelity aerodynamic inference of non-linear flow fields without requiring CFD or meshing expertise, supporting both surface- and volume-based surrogate training, real-time inference, and exploration of shape–performance trade-offs.
The current release focuses on the high-speed cruise configuration with only fuselage and wing, emphasizing planform design. A fully parametric CRM was constructed in OnShape by importing the NASA reference, deconstructing it, and reassembling it with exposed variables. The parameterization separates wing and fuselage and proceeds as follows:
Wing: intersect the reference wing at six spanwise stations to extract airfoil profiles; expose parameters controlling profile translation and rotation (twist); re-loft the wing and reconstruct the tip geometry.
Fuselage: parameterize fuselage length and radius; size the wing–body fairing from the local chord of the newly lofted wing.
Assembly: combine fuselage and wing via Boolean operations. While many micro-parameters are used internally, the dataset varies a compact set of macro planform/fuselage parameters (seven in the current release), with micro-parameters derived from these.
Datasets are generated in batches at fixed Mach numbers; geometry parameters and angle of attack are selected via Latin hypercube sampling to span distinct flow regimes intentionally. Lower-Mach batches avoid shocks, whereas higher-Mach batches exhibit complex three-dimensional shock structures, enabling targeted evaluation of AI/ML methods across qualitatively different physics. The SHIFT-Wing design space targets classical wing planform and fuselage variables around the NASA CRM reference.
The design and operating parameters are varied as follows. Additional details may be found in the dataset provided by Luminary et al. [14].
On the fuselage, aspect ratio varies from 7.5 to 11 with a reference value of 9. The quarter-chord sweep angle spans 25 to 37.5 (reference 35). The root-chord extension factor ranges from 1.0 to 1.4 (reference 1.373). Fuselage diameter ranges from 240 to 258 inches, with 240 inches as the reference.
On the wing, root twist varies from 3 to 9 (reference 6.717). Spanwise twist is controlled via deltas: root-to-break ranges from −7 to −3 (reference −5.953), and break to tip ranges from −7.5 to −1.5 (reference −4.513).
Operating conditions are varied through angle of attack from 0 to 4 and Mach number from 0.5 to 0.85.
All simulations are run on the Luminary Cloud platform using a GPU-native finite-volume solver. For SHIFT-Wing, turbulence is modeled via steady RANS with the Spalart–Allmaras model [57]. All vehicle surfaces are treated as no-slip walls, and angle of attack is imposed through far-field boundary conditions. A key distinction from SHIFT-SUV is meshing: SHIFT-Wing employs Luminary Mesh Adaptation (LMA), a proprietary solution-adaptive meshing workflow that iteratively refines local mesh density and anisotropy to capture sharp features (e.g., transonic shocks) without user intervention. This automated sequence of meshes and solutions improves accuracy across diverse geometries and flow conditions while eliminating manual meshing effort.
For training, we use 1698 simulations (1138 at Mach 0.5 and 560 at Mach 0.85) from this dataset split randomly into 80/10/10. We match the exact simulations in each split with the AB-UPT benchmarking work in Alkin et al. [47].
Crash-dynamics datasets
As an additional example beyond the aerodynamics datasets, we evaluate GeoTransolver on two nonlinear crash-dynamics benchmarks: a bumper-beam impact dataset and a full-vehicle Body-in-White (BIW) frontal-impact dataset. These datasets test prediction of transient structural response under large deformation, contact, and elastoplastic effects.
The bumper-beam dataset is generated from a parametric finite-element representation of a front-impact bumper-beam scenario. The bumper beam is modeled with shell elements and impacted by a cylindrical rigid wall along the longitudinal direction. The design space varies geometric scaling, impact velocity, shell thickness, and transverse impactor offset, resulting in 135 simulations. The dataset is split into 80% training, 10% validation, and 10% testing.
The BIW crash dataset is generated from a high-fidelity finite-element model of a vehicle Body-in-White undergoing frontal impact against a rigid barrier at 56 km/h. The model contains approximately 400,000 nodes and 380,000 elements. The design space is constructed by perturbing the thicknesses of 33 front-end structural components by ±20% around their nominal values, resulting in 150 simulations. Each simulation spans 120 ms and contains 25 discrete time steps. The dataset is split into 90% training, 5% validation, and 5% testing.
For the BIW benchmark, we additionally evaluate position, velocity, and acceleration histories at driver and passenger toe-pan probe locations, which are sensitive to front-end deformation and are relevant to crashworthiness assessment.
Models
We evaluate GeoTransolver on SHIFT-SUV and SHIFT-Wing alongside strong baselines—Transolver and DοMINO—and report AB-UPT results directly from Alkin et al. [47]. To probe architectural sensitivity, we conduct an ablation study on DrivAerML. Transolver uses the implementation released in PhysicsNeMo, with full self-attention over concatenated surface and volume tokens (10M parameters). DοMINO is run with the default PhysicsNeMo configuration following Ranade et al. [8], including multi-scale neighborhoods and auxiliary geometric features (surface normals, areas, signed distance), yielding 19.7M parameters. AB-UPT is not re-implemented or re-run; we adopt the refined 384-dimensional results from Alkin et al. [47]. GeoTransolver, released in PhysicsNeMo, spans 10–25M parameters depending on GALE depth, ball-query radii, and kernel size; § 5.1 reports accuracy across these architectural choices on DrivAerML. For SHIFT-Wing, GeoTransolver conditions each block on global parameters (angle of attack, Mach) via geometry/global context projections, in contrast to Transolver’s plain token conditioning. All models are trained for up to 500 epochs on a single NVIDIA GB200 node using the Muon optimizer [59], under a shared preprocessing and evaluation protocol.
For the crash-dynamics benchmarks, we compare Transolver, GeoTransolver, and GeoTransolver with a FLARE attention backend under matched training protocols. Transolver and GeoTransolver use 128 latent tokens for physics attention, while the FLARE-backend variant uses 128 global queries for low-rank attention. For the bumper-beam dataset, GeoTransolver uses 6 layers, 8 attention heads, 256 hidden channels, and multi-scale ball-query radii of [0.05, 0.25] with [8, 32] neighbors. For the BIW crash dataset, GeoTransolver uses 5 layers, 8 attention heads, and the same multi-scale ball-query configuration. In the FLARE-backend variant, the geometry-aware context pathway is kept fixed and only the self-attention/token-mixing backend is replaced.
Metrics
The prediction accuracy is compared across the metrics mentioned below and averaged over test samples. refer to the index of the test sample and denotes the model prediction.
Mean Absolute Error:
Relative Norm:
For surface predictions, we also compare lift and drag forces on the test geometries. These global metrics are commonly used during design exploration for industrial engineering applications. The total force vector on test geometries is obtained as follows:
where, denotes the surface pressure, is the shear stress, is the normal vector. For calculating the drag force, the normal vector is set in the tangential direction, , while for lift force it is in the direction perpendicular to flow, . The lift and drag forces are computed for each of the test cases using both predicted and ground truth fields and the value is calculated over the entire test set, to determine the predictive accuracy.
For the best-performing GeoTransolver configuration, we provide qualitative assessments: contour plots of key fields (e.g., pressure coefficient, velocity magnitude) and design trends showing predicted drag/lift as functions of representative geometry or regime parameters, alongside ground-truth curves where available.
For crash dynamics, we report the relative error over the predicted spatiotemporal structural response,
where and denote the ground-truth and predicted structural states for test sample . For the BIW benchmark, we also report mean squared error for position, velocity, and acceleration at driver and passenger toe-pan probe locations.
Computational cost of multi-scale ball queries
Table 5 reports per-step training and inference throughput for each ball-query configuration on DrivAerML for both surface and volume tasks (GALE backend, 8×H100, batch size 8, bfloat16, torch.compile). Mean ± std are reported across seeds where available; single-seed entries report mean and std over within-epoch step times. Single-scale ball queries add 1.1–1.2× overhead on surface; the 4- and 6-scale multi-radii configurations add approximately 2× but plateau between them. The 6-scale entry required reducing query points from 200k to 160k to fit within H100’s 80 GB memory.
*160k query points (other configs use 200k) to fit in GPU memory.
Query and geometry token ablation
Table 6 reports relative errors on DrivAerML as a function of query and geometry token counts. Increasing geometry token density at fixed queries sharply reduces volume errors (e.g., at 20k queries, drops from 11.1% at 50k geo to 3.09% at 300k), highlighting the need for adequate geometric coverage. Scaling query tokens improves surface metrics, with the best overall surface/integral performance at 60k/300k. Volume pressure exhibits a sweet spot at 60k/150k, whereas very high geometry counts with moderate queries can degrade volume errors, suggesting imbalance.
| Query/Geo | Surface | Surface | Surface | Surface | Volume | Volume |
| 20k/50k | ||||||
| 20k/150k | ||||||
| 20k/300k | ||||||
| 40k/50k | ||||||
| 40k/150k | ||||||
| 40k/300k | ||||||
| 60k/50k | ||||||
| 60k/150k | ||||||
| 60k/300k |
Table 6. Relative errors (%) on DrivAerML test geometries for number of geometry and query points.
Additional results
In this section, we provide additional results for the three aerodynamics datasets evaluated in this paper, DrivAerML, SHIFT-SUV, and SHIFT-Wing, as well as the two crash-dynamics benchmarks. For the aerodynamics datasets, we compare GeoTransolver predictions with simulated ground truth to assess its ability to capture design trends and surface and volume contours. For crash dynamics, we assess deformation fields and safety-relevant probe responses.
DrivAerML
Figure 2 shows drag and lift coefficients across the 48 test designs, with designs ordered by ascending ground-truth values. The predicted curves closely track the ground-truth trends, preserving the overall ordering and capturing large-scale variations. Small oscillations appear between adjacent designs with subtle directional changes, indicating minor non-monotonic deviations in tightly clustered regions.

Figure 2. Design trends: drag and lift vs. test designs.
The design points with the highest and lowest drag and lift forces correspond to out-of-distribution and are never seen during training. It may be noted that the model captures the aerodynamic forces for these designs reasonably accurately and the predictions can be further improved with augmentation of the training set.
Surface-field SHIFT results
Table 7 reports the surface-field counterpart to the volume table in the main text (Table 3). GeoTransolver leads on most surface metrics with two exceptions: surface pressure at Mach 0.85 (AB-UPT) and wall-shear stress at Mach 0.5 (DoMINO, trained per-Mach).
| Model | Estate | Estate | Fastback | Fastback | Ma=0.5 | Ma=0.5 | Ma=0.85 | Ma=0.85 |
| GeoTransolver | 0.0057 | 3.81 | 0.0056 | 3.70 | 0.021 | 12.2 | 0.081 | 13.01 |
| AB-UPT | 0.0064 | 4.95 | 0.0064 | 5.03 | 0.022 | 12.5 | 0.079 | 13.3 |
| DoMINO | 0.0100 | 12.24 | 0.0100 | 11.74 | 0.468 | 10.2 | 1.88 | 13.35 |
| Transolver | 0.0079 | 4.98 | 0.0078 | 4.97 | 0.094 | 12.4 | 0.098 | 13.2 |
Table 7. Surface-field relative errors (%) across SHIFT-SUV (Estate, Fastback) and SHIFT-Wing (Ma=0.5, Ma=0.85). DoMINO is trained per-Mach on SHIFT-Wing; other models share the combined dataset.
SHIFT-SUV
Figure 3 and Figure 4 plot drag and lift coefficients across the test design IDs for the Estate and Fastback. The predicted trends closely follow ground truth, capturing overall ordering and directional changes. Larger oscillations are visible in Estate lift, likely due to sparser coverage in that lift regime.
Figure 3. Design trends for Estate: drag and lift vs. test designs.
Figure 4. Design trends for Fastback: drag and lift vs. test designs.
Figure 5 and Figure 6 compare surface pressure and wall-shear stress contours between GeoTransolver and ground truth. Errors are uniformly distributed and bounded, indicating high accuracy; the largest discrepancies occur near wheels and mirrors, where flow is particularly complex.
Figure 5. Surface contour comparisons for Estate.
Figure 6. Surface contour comparisons for Fastback.
Similarly, Figure 7 and Figure 8 show volume comparisons on the plane. The highest errors appear near separation regions downstream and in the wake where gradients are sharp, but overall predictions remain accurate.
Figure 7. Volume contour comparisons for Estate on the plane.
Figure 8. Volume contour comparisons for Fastback on the plane.
SHIFT-Wing
Figure 9 and Figure 10 plot drag and lift across test design IDs at Mach 0.5 and 0.85. Predicted trends closely track ground truth in both regimes, preserving ordering and directional changes.
Figure 9. Design trends for Mach 0.5: drag and lift vs. test designs.
Figure 10. Design trends for Mach 0.85: drag and lift vs. test designs.
Figure 11 and Figure 12 compare surface pressure and wall-shear stress contours between GeoTransolver and ground truth. Errors are uniformly distributed and bounded, indicating high accuracy; the largest discrepancies appear on the wing, especially at higher Mach where sharper transitions occur.
Figure 11. Surface contour comparisons for Mach 0.5.
Figure 12. Surface contour comparisons for Mach 0.85.
Figures 13 and 14 show volume comparisons on the plane. Errors are most pronounced over the wing and fuselage and in the downstream wake where separation and steep gradients occur, particularly at Mach 0.85. Overall, predictions remain accurate across both regimes.
Crash dynamics
In this section, we provide additional results for the bumper-beam and BIW crash-dynamics datasets. We compare GeoTransolver predictions with finite-element ground truth to assess its ability to capture final deformation states, deformation evolution over time, and safety-relevant probe responses.
| Model | Training speed | No. of parameters |
| Transolver | 4.8 s/epoch | 3,700,783 |
| GeoTransolver | 7.5 s/epoch | 5,686,504 |
| GeoTransolver with FLARE | 7.1 s/epoch | 5,150,624 |
Table 8. Training speed and number of parameters for the BIW crash-dynamics benchmark using the Muon optimizer.
Table 9 shows that GeoTransolver improves the prediction of position, velocity, and acceleration at both driver and passenger probe locations, and that replacing the physics-attention backend with FLARE further improves these safety-relevant quantities. These probe-level results complement the full-field relative errors reported in the main text and indicate that the geometry-aware context remains effective when paired with an alternative efficient attention backend.
| Probe | Model | Position | Velocity | Acceleration |
| Driver | Transolver | |||
| Driver | GeoTransolver | |||
| Driver | GeoTransolver with FLARE | |||
| Passenger | Transolver | |||
| Passenger | GeoTransolver | |||
| Passenger | GeoTransolver with FLARE |
Table 9. Probe-level MSE comparison for BIW crash dynamics using the Muon optimizer. Position, velocity, and acceleration are evaluated at driver and passenger toe-pan locations.
Figure 15, Figure 16, and Figure 17 provide qualitative deformation comparisons for GeoTransolver with the FLARE attention backend on the crash-dynamics benchmarks. These visual results help localize prediction differences during large deformation and contact-dominated response.

Figure 15. Visual comparison of BIW crash deformation for representative test samples. The top row shows finite-element ground truth, and the bottom row shows GeoTransolver predictions using the FLARE attention backend.

Figure 16. (a) BIW crash deformation evolution for the first held-out test sample. (b) BIW crash deformation evolution for the second held-out test sample. Temporal evolution of BIW crash deformation for the same held-out test samples shown in Figure 15. Columns show representative time steps through the impact event. For each sample, the top row shows finite-element ground truth and the bottom row shows GeoTransolver predictions using the FLARE attention backend.

Figure 17. Visual comparison of bumper-beam crash deformation for representative test samples. The top row shows finite-element ground truth, and the bottom row shows GeoTransolver predictions using the FLARE attention backend.
Figure 13. Volume contour comparisons on the plane for Mach 0.5.
Figure 14. Volume contour comparisons on the plane for Mach 0.85.
References
- [1]Xu-Wen Zhao, Xiao-Meng Tong, Fang-Wei Ning, Mao-Lin Cai, Fei Han, and Hong-Guang Li. Review of empowering computer-aided engineering with artificial intelligence. Advances in Manufacturing, pages 1–41, 2025.DOI
- [2]Sofos Filippos. Advancing fluid mechanics with artificial intelligence and machine learning. Fluids, 10 (11):297, 2025.DOI
- [3]Steven L Brunton, Bernd R Noack, and Petros Koumoutsakos. Machine learning for fluid mechanics. Annual review of fluid mechanics, 52(1):477–508, 2020.
- [4]Pushan Sharma, Wai Tong Chung, Bassem Akoush, and Matthias Ihme. A review of physics-informed machine learning in fluid mechanics. Energies, 16(5):2343, 2023.DOI
- [5]Neil Ashton, Johannes Brandstetter, and Siddhartha Mishra. Fluid intelligence: A forward look on ai foundation models in computational fluid dynamics. arXiv preprint arXiv:2511.20455, 2025.arxiv.org/abs/2511.20455
- [6]Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for pdes on general geometries. arXiv preprint arXiv:2402.02366, 2024.arxiv.org/abs/2402.02366
- [7]Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015.arxiv.org/abs/1505.04597
- [8]Rishikesh Ranade, Mohammad Amin Nabian, Kaustubh Tangsali, Alexey Kamenev, Oliver Hennig, 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.arxiv.org/abs/2501.13350
- [9]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. arXiv preprint arXiv:2502.09692, 2025.arxiv.org/abs/2502.09692
- [10]Shizheng Wen, Arsh Kumbhat, Levi Lingsch, Sepehr Mousavi, Yizhou Zhao, Praveen Chandrashekar, and Siddhartha Mishra. Geometry aware operator transformer as an efficient and accurate neural surrogate for pdes on arbitrary domains. 2025.arxiv.org/abs/2505.18781
- [11]NVIDIA. NVIDIA PhysicsNeMo, 2025. URL https://developer.nvidia.com/physicsnemo.
- [12]Neil Ashton, Charles Mockett, Marian Fuchs, Louis Fliessbach, Hendrik Hetmann, Thilo Knacke, Norbert Schonwald, Vangelis Skaperdas, Grigoris Fotiadis, Astrid Walle, et al. Driverml: High-fidelity computational fluid dynamics dataset for road-car external aerodynamics. arXiv preprint arXiv:2408.11969, 2024.arxiv.org/abs/2408.11969
- [13]Luminary Cloud. Shift-suv: High-fidelity computational fluid dynamics dataset for suv external aerodynamics. URL https://huggingface.co/datasets/luminary-shift/SUV/, 2025.
- [14]Luminary Cloud. Shift-wing: High-fidelity computational fluid dynamics dataset for transonic wing external aerodynamics. URL https://huggingface.co/datasets/luminary-shift/WING/, 2025.
- [15]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
- [16]Vedant Puri, Aditya Joglekar, Sri Datta Ganesh Bandreddi, Kevin Ferguson, Yu-hsuan Chen, Yongjie Jessica Zhang, and Levent Burak Kara. Flare: Fast low-rank attention routing engine. arXiv preprint arXiv:2508.12594, 2025.arxiv.org/abs/2508.12594
- [17]Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24(89):1–97, 2023.
- [18]Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218–229, 2021.
- [19]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895, 2020.arxiv.org/abs/2010.08895
- [20]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Andrew Stuart, Kaushik Bhattacharya, and Anima Anandkumar. Multiple graph neural operator for parametric partial differential equations. Advances in Neural Information Processing Systems, 33:6755–6766, 2020.arxiv.org/abs/2006.09535
- [21]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial differential equations. arXiv preprint arXiv:2003.03485, 2020.arxiv.org/abs/2003.03485
- [22]Zongyi Li, Daniel Zhengyu Huang, Burigede Liu, and Anima Anandkumar. Fourier neural operator with learned deformations for pdes on general geometries. Journal of Machine Learning Research, 24(388):1–26, 2023.DOI
- [23]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.arxiv.org/abs/1706.03762
- [24]Zhongkai Hao, Zhengyi Wang, Hang Su, Chengyang Ying, Yinpeng Dong, Songming Liu, Ze Cheng, Jian Song, and Jun Zhu. Gnot: A general neural operator transformer for operator learning. In International Conference on Machine Learning, pages 12556–12569. PMLR, 2023.arxiv.org/abs/2302.14376
- [25]Andrey Bryutkin, Jiahao Huang, Zhongying Deng, Guang Yang, Carola-Bibiane Schönlieb, and Angelica Aviles-Rivero. Hamlet: Graph transformer neural operator for partial differential equations. arXiv preprint arXiv:2402.03541, 2024.arxiv.org/abs/2402.03541
- [26]Benjamin Shih, Ahmad Peyvan, Zhongqiang Zhang, and George Em Karniadakis. Transformers as neural operators for solutions of differential equations with finite regularity. Computer Methods in Applied Mechanics and Engineering, 434:117560, 2025.arxiv.org/abs/2405.19166
- [27]Qibang Liu, Weiheng Zhong, Hadi Meidani, Diab Abueidda, Seid Koric, and Philippe Geubelle. Geometry-informed neural operator transformer. arXiv preprint arXiv:2504.19452, 2025.arxiv.org/abs/2504.19452
- [28]Oded Ovadia, Adar Kahana, Panos Stinis, Eli Turkel, Dan Givoli, and George Em Karniadakis. Vito: Vision transformer-operator. Computer Methods in Applied Mechanics and Engineering, 428:117109, 2024.arxiv.org/abs/2303.08891
- [29]Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh-based simulation with graph networks. In International conference on learning representations, 2020.arxiv.org/abs/2010.03409
- [30]Ugo Pelissier, Augustin Parret-Fréaud, Felipe Bordeu, and Youssef Mesri. Graph neural networks for mesh generation and adaptation in structural and fluid mechanics. Mathematics, 12(18):2933, 2024.DOI
- [31]Pourya Pilva and Ahmad Zareei. Learning time-dependent pde solver using message passing graph neural networks. arXiv preprint arXiv:2204.07651, 2022.arxiv.org/abs/2204.07651
- [32]Rini Jasmine Gladstone, Helia Rahmani, Vishvas Suryakumar, Hadi Meidani, Marta D’Elia, and Ahmad Zareei. Mesh-based gnn surrogates for time-independent pdes. Scientific reports, 14(1):3394, 2024.DOI
- [33]Mohammad Amin Nabian, Chang Liu, Rishikesh Ranade, and Sanjay Choudhury. X-meshgraphnet: Scalable multi-scale graph neural networks for physics simulation. arXiv preprint arXiv:2411.17164, 2024.arxiv.org/abs/2411.17164
- [34]Meire Fortunato, Tobias Pfaff, Peter Wirnsberger, Alexander Pritzel, and Peter Battaglia. Multiscale meshgraphnets. arXiv preprint arXiv:2210.00612, 2022.arxiv.org/abs/2210.00612
- [35]Chris Choy, Alexey Kamenev, Jean Kossaifi, Max Rietmann, Jan Kautz, and Kamyar Azizadnesheli. Factorized implicit global convolution for automotive computational fluid dynamics prediction. arXiv preprint arXiv:2502.04317, 2025.arxiv.org/abs/2502.04317
- [36]Qian Chen, Mohamed Elfaraie, Angela Dai, and Faez Ahmed. Tripnet: Learning large-scale high-fidelity 3d car aerodynamics with triplane networks. arXiv preprint arXiv:2503.17400, 2025.DOI
- [37]Moshe Eliasof, Eldad Haber, and Eran Treister. Pde-gcn: Novel architectures for graph neural networks motivated by partial differential equations. Advances in neural information processing systems, 34:3836–3849, 2021.arxiv.org/abs/2108.01938
- [38]Rishikesh Ranade, Chris Hill, Lalit Ghule, and Jay Pathak. A composable machine-learning approach for steady-state simulations on high-resolution grids. Advances in Neural Information Processing Systems, 35:17386–17401, 2022.arxiv.org/abs/2210.05837
- [39]Wenzhuo Liu, Mouadh Yagoubi, and Marc Schoenauer. Multi-resolution graph neural networks for pde approximation. In International Conference on Artificial Neural Networks, pages 151–163. Springer, 2021.DOI
- [40]Yuying Liu, Colin Ponce, Steven L Brunton, and J Nathan Kutz. Multiresolution convolutional autoencoders. Journal of Computational Physics, 474:118101, 2023.arxiv.org/abs/2004.04946
- [41]John Guibas, Morteza Mardani, Zongyi Li, Andrew Tao, Anima Anandkumar, and Bryan Catanzaro. Adaptive fourier neural operators: Efficient token mixers for transformers. arXiv preprint arXiv:2111.13587, 2021.arxiv.org/abs/2111.13587
- [42]Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019.DOI
- [43]Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Aziz-Azadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations. ACM/IMS Journal of Data Science, 1(3):1–27, 2024.arxiv.org/abs/2111.03794
- [44]Victor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In International conference on machine learning, pages 9323–9332. PMLR, 2021.
- [45]Johannes Brandstetter, Rob Hesselink, Elise van der Pol, Erik J Bekkers, and Max Welling. Geometric and physical quantities improve e (3) equivariant message passing. arXiv preprint arXiv:2110.02905, 2021.arxiv.org/abs/2110.02905
- [46]Yi-Lun Liao and Tess Smidt. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. arXiv preprint arXiv:2206.11990, 2022.
- [47]Benedikt Alkin, Richard Kurle, Louis Serrano, Dennis Just, and Johannes Brandstetter. Ab-upt for automotive and aerospace applications. arXiv preprint arXiv:2510.15808, 2025.arxiv.org/abs/2510.15808
- [48]Philippe R Spalart, Shur Deck, Michael L Shur, Kyle D Squires, M Kh Strelets, and Andrei Travin. A new version of detached-eddy simulation, resistant to ambiguous grid densities. Theoretical and computational fluid dynamics, 20(3):181–195, 2006.DOI
- [49]Bruno Chaouat. The state of the art of hybrid rans/les modeling for the simulation of turbulent flows. Flow, turbulence and combustion, 99(2):279–327, 2017.DOI
- [50]Stefan Heinz. A review of hybrid rans-les methods for turbulent flows: Concepts and applications. Progress in Aerospace Sciences, 114:100597, 2020.DOI
- [51]Neil Ashton, Paul Batten, Andrew W Cary, Kevin R Holst, and Vangelis Skaperdas. Hlpw-4/gmgw-3: Hybrid rans/les technology focus group workshop summary. In AIAA Aviation 2022 Forum, page 3293, 2022.DOI
- [52]Burkhard Hupertz, Neil Lewington, Charles Mockett, Neil Ashton, and Lian Duan. Towards a standardized assessment of automotive aerodynamic cfd prediction capability-autocfd 2: Ford drivaer test case summary. Technical report, SAE Technical Paper, 2022.DOI
- [53]Chenyi Zhang, Max Tanneberger, Timo Kuthada, Felix Wittmeier, Jochen Wiedemann, and Juliane Nies. Introduction of the aerosuv-a new generic suv model for aerodynamic research. Technical report, SAE Technical Paper, 2019.DOI
- [54]Angelina I Heft, Thomas Indinger, and Nikolaus A Adams. Experimental and numerical investigation of the drivaer model. In Fluids engineering division summer meeting, volume 44755, pages 41–51. American Society of Mechanical Engineers, 2012.DOI
- [55]Chenyi Zhang. Aerodynamic study on the vehicle shape parameters with respect to ground simulation. Springer, 2021.
- [56]T Economon D Engwirda J Ho M Mara and R Mansharamani. Luminary contribution to the autocfd4 workshop. URL https://autocfd.org/autocfd4/, 2024.
- [57]Joshua Krakos, Pedro Gomes, Jonathan Ho, Michael Mara, Gonzalo Sáez, Adrien Loseille, Thomas Economon, and Juan J Alonso. Gpu-based and adaptive solution technology for the 5th aiaa high lift prediction workshop. In AIAA SCITECH 2025 Forum, page 0496, 2025.DOI
- [58]Nasa common research model project. high speed common research model (crm). https://commonresearchmodel.larc.nasa.gov/home-2/high-speed-crm/.
- [59]Chongjie Si, Debing Zhang, and Wei Shen. Adamuon: Adaptive muon optimizer. arXiv preprint arXiv:2507.11005, 2025.arxiv.org/abs/2507.11005