Command Palette

Search for a command to run...

Architectures

A quick overview of what each architecture tends to be good at, based on how they show up on our current benchmarks.

Geometry-aware transformer

GeoTransolver

Transformer with multi-scale geometry injected at every layer. Strong on large surface/volume CFD and some transient structural cases; multi-scale context adds compute cost.

Strengths

  • Persistent geometry conditioning across depth
  • Multi-scale neighborhoods for non-uniform engineering meshes
  • Surface, volume, and transient structural examples in one framework

Limitations & caveats

  • Published accuracy is strongest on large accelerator configurations
  • Multi-scale context can roughly double throughput cost
  • Crash evidence is limited to short horizons and a small number of benchmark families
Compute
Multi-scale ball queries add material compute and memory overhead. Our DrivAerML reference run trained on one grossular-1 node (1× NVIDIA A10).
Preprocessing
Requires geometry point clouds, neighborhood queries, and task-specific point features.

Submissions

ProblemSubmittedRun
ROEB-Aero: DrivAerMLJul 22, 202652bd9
ROEB-Crash: OpenRadioss CarJul 8, 2026984c9

Physics-attention transformer

Transolver

Transformer that learns physics tokens over mesh or grid points without explicit multi-scale geometry injection. Often faster to train than GeoTransolver; less geometry context at depth.

Strengths

  • Strong baseline on moderate-size CFD meshes (airfoil, plate)
  • Attention over learned physics tokens scales to many point-cloud formulations
  • Lower preprocessing burden than full geometry-token pipelines

Limitations & caveats

  • Less explicit multi-scale geometry context than GeoTransolver
  • Native multi-million-node industrial meshes may need capacity tuning
  • Not a drop-in substitute for GeoTransolver on every benchmark row
Compute
Typical Rescale grossular-1 runs finish in tens of minutes to a few hours depending on benchmark and epochs.
Preprocessing
Point or mesh features plus normalization statistics from the training split.

Submissions

ProblemSubmittedRun
ROEB-Aero: 2D AirfoilJun 24, 202668434
ROEB-Stress: Plate with HoleJun 24, 20269b52e

Geometry-aware transformer

Two-Stream Transformer

Dual-stream transformer for steady CFD on surface or volume fields. Fits airfoil and plate-scale problems; not aimed at FE transient crash sequences.

Strengths

  • Strong on 2D/ moderate 3D steady CFD with clear global scalars
  • Handles native open meshes on several benchmarks without heavy decimation
  • Published Rescale AI Physics recipes for airfoil and plate

Limitations & caveats

  • Targets steady single-frame fields, not transient structural time series
  • Not applicable on OpenRadioss Car (fea-deform transient)
  • Very large volume meshes may need the same capacity planning as other transformers
Compute
Typical grossular-1 training runs from tens of minutes to a few hours.
Preprocessing
Mesh or point features, case parameters, and normalization from the training split.

Submissions

ProblemSubmittedRun
ROEB-Aero: DrivAerMLJul 20, 2026e1062
ROEB-Stress: Plate with HoleJun 18, 202634780
ROEB-Aero: 2D AirfoilJun 17, 20267485a

Multi-scale neural operator

DoMINO

Decomposable point-cloud operator for large surface and volume CFD. Good when you can invest in SDF-style preprocessing; surface and volume models are often trained separately.

Strengths

  • Local evaluation can avoid processing every output point simultaneously
  • Multi-scale geometry encoding captures near and far context
  • Public surface and volume recipes are available

Limitations & caveats

  • Preprocessing is substantial and dataset-specific
  • Surface and volume models are commonly trained separately
  • Very large volume cases create significant storage and I/O pressure
Compute
Heavy preprocessing and long wall times are common. Our DrivAerML reference run used one grossular-1 node (1× NVIDIA A10) for ~31 hours.
Preprocessing
Signed-distance fields, nearest-neighbor features, bounding boxes, normalization statistics, and optional cached geometry encodings.

Submissions

ProblemSubmittedRun
ROEB-Aero: DrivAerMLJul 21, 2026569e9
ROEB-Stress: Plate with HoleJun 18, 20261e4d5
ROEB-Aero: 2D AirfoilJun 17, 20264dfb6

Graph neural network

MeshGraphNet

Message passing over mesh connectivity. Natural for transient rollout on a fixed topology; full industrial graphs often need partitioning or decimation.

Strengths

  • Mesh connectivity provides a strong local physical prior
  • Natural fit for transient rollout on a consistent topology
  • Mature message-passing ecosystem and multiple scalable variants

Limitations & caveats

  • Long-range effects require depth, hierarchy, or multi-scale variants
  • Full industrial meshes can exceed practical accelerator memory
  • Scalable variants should not be attributed to the base architecture without qualification
Compute
Memory grows with graph size and edge count; industrial cases often require partitioning, coarsening, or decimation.
Preprocessing
Graph construction, edge features, consistent topology handling, and often mesh partitioning.

Submissions

ProblemSubmittedRun
ROEB-Crash: OpenRadioss CarJul 13, 202684260
ROEB-Stress: Plate with HoleJun 18, 20265187f
ROEB-Aero: 2D AirfoilJun 17, 20265d7f7

Attention-based surrogate

Transformer (general)

Attention over tokens derived from mesh points, patches, or physics-aware embeddings. Flexible, but geometry and scale choices dominate whether it works on industrial meshes.

Strengths

  • Flexible tokenization for many mesh and parameter layouts
  • Strong when global context and multi-field outputs matter
  • Large ecosystem of training tricks and scaling recipes

Limitations & caveats

  • Quadratic attention cost without specialized variants
  • Needs careful geometry encoding for unstructured industrial meshes
  • Compute and memory scale quickly on million-node surfaces

Submissions

No submissions on this architecture yet.

Graph neural network

Graph neural network (general)

Message passing on mesh or k-NN graphs. Excellent local physics prior; watch memory on dense industrial connectivity.

Strengths

  • Natural fit for mesh adjacency and local conservation-like patterns
  • Well suited to transient steps on fixed topology
  • Many published variants (MeshGraphNet, GraphCast-style stacks)

Limitations & caveats

  • Long-range coupling needs depth or hierarchical graphs
  • Edge count drives memory on fine industrial meshes
  • Decimated graphs change the problem — compare like with like

Submissions

No submissions on this architecture yet.

Neural operator

Neural operator (general)

Learns mappings between function spaces (FNO, DeepONet, GNO, etc.). Strong for parametric PDE families when training distribution matches deployment.

Strengths

  • Designed for family-of-PDEs and parameter sweeps
  • Can generalize across boundary conditions when trained broadly
  • Multiple public operator architectures to build on

Limitations & caveats

  • Often assumes structured grids or specific query patterns
  • Unstructured million-node CFD may need hybrid or point-cloud operators (e.g. DoMINO-style)
  • Preprocessing and normalization dominate transfer quality

Submissions

No submissions on this architecture yet.

Convolutional network

CNN (general)

Convolution on structured grids or voxelized fields. Fast and mature when the problem can be cast to a regular lattice.

Strengths

  • Efficient on structured or voxelized representations
  • Mature tooling and predictable memory
  • Good baseline for 2D fields and parametric sweeps on grids

Limitations & caveats

  • Unstructured industrial meshes need resampling or projection
  • Geometry fidelity can be lost in voxelization
  • Less natural for native CAD surface meshes at full resolution

Submissions

No submissions on this architecture yet.

Hybrid architecture

Hybrid (general)

Combines operators (e.g. GNN + transformer, FNO + CNN). Useful when no single inductive bias covers the whole mesh — document which parts do what.

Strengths

  • Can mix local mesh bias with global attention or operators
  • Room to specialize submodules for parameters vs fields

Limitations & caveats

  • More moving parts — harder to compare fairly across submissions
  • Preprocessing and compute cost sum across branches
  • Reviewers need clarity on which components are novel vs off-the-shelf

Submissions

No submissions on this architecture yet.

Benchmark fit (quick reference)

BenchmarkCommon architecturesWatch out for
Plate with holeTwo-Stream Transformer, MeshGraphNet, Transolver, DoMINOModerate mesh — most families fit; compare train time fairly.
2D AirfoilTwo-Stream Transformer, MeshGraphNet, Transolver, DoMINOSurface Cp + global Cl/Cd — good 2D steady CFD testbed.
DrivAerMLDoMINO, Two-Stream Transformer, GeoTransolverNative ~8M-node surfaces; vanilla MeshGraphNet needs heavy decimation (different problem).
OpenRadioss CarMeshGraphNet, GeoTransolverFE transient displacement — TST and DoMINO do not apply.

See also the FAQ on DrivAerML and MeshGraphNet and the evaluation protocol.