Blog
RHX BLOG

Open CFD Datasets Are Changing AI Aerodynamics

June 21, 20261 min read

CFDAI CAEDatasetAerodynamics리서치
CFD
Open CFD Datasets Are Changing AI Aerodynamics

Visual module

Visual review map

A compact map of the article: decision, input, validation, and output.

decision-ready evidence
01

Episode

CAD가 아니라 해석 가능한 사건

02

Schema

재료, load case, solver provenance

03

Validation

시험값, uncertainty, OOD split

04

Learning

모델이 배울 수 있는 결정 맥락

One of the biggest bottlenecks in AI aerodynamics has been data quality. Industrial CFD depends on geometry, mesh, turbulence model, boundary conditions, solver setup, and post-processing. Without transparent datasets, AI models stay trapped in toy problems.

Important datasets

DrivAerNet provides thousands of car meshes and aerodynamic data. DrivAerNet++ expands this to 8,000 diverse car designs and more than 39 TB of public engineering data. AhmedML contributes 500 Ahmed body variants with high-fidelity CFD. WindsorML provides 355 Windsor body variants using high-fidelity WMLES with very large cell counts.

Benchmarking matters

PhysicsNeMo's DoMINO example uses DrivAerML, and recent benchmarking work compares DoMINO, X-MeshGraphNet, and FIGConvNet on automotive aerodynamics metrics. This moves AI CAE from isolated model claims to comparable engineering evidence.

What to inspect

  • Geometry diversity.
  • Solver fidelity and turbulence model.
  • Mesh, domain, and boundary-condition consistency.
  • Available outputs: coefficients, surface fields, volume fields.
  • Validation, license, and reproducibility.

References

Open CFD Datasets Are Changing AI Aerodynamics | RHX.LAB