DrivAerML Validation Study
Gradient Dynamics benchmarked its GPU-native steady RANS solver against the DrivAerML public automotive aerodynamics dataset to evaluate accuracy, mesh convergence, robustness across vehicle variants, and workflow readiness for industrial design studies.
Study Overview
DrivAerML extends the DrivAer notchback benchmark with 500 parametrically morphed vehicle geometries and high-fidelity reference CFD data. For this validation study, Gradient Dynamics evaluated three representative variants: Run 1, Run 166, and Run 402.
Simulations used the Gradient Dynamics GPU-native steady Reynolds-Averaged Navier-Stokes solver with the k-omega SST turbulence model. The case setup used a 60 m/s inlet velocity, rolling-road ground condition, and drag coefficient as the primary validation metric.
Drag Coefficient Validation
Across the three validation cases, the Gradient Dynamics solver reproduced the DrivAerML reference drag coefficients within 3.0%, with two cases within approximately 1.1% of the published values.
| Case | Gradient Dynamics Cd | DrivAerML Reference Cd | Difference |
|---|---|---|---|
| Run 1 | 0.295 | 0.304 | -3.0% |
| Run 166 | 0.247 | 0.245 | +0.8% |
| Run 402 | 0.261 | 0.264 | -1.1% |
Mesh Sensitivity
A mesh-sensitivity study on Run 1 showed consistent convergence toward the DrivAerML reference value as cell count, near-wall resolution, and boundary-layer resolution increased. The fine mesh reached a drag coefficient error of -0.49% relative to the reference result.
| Mesh Level | Cell Count | Target y+ | Layers | Cd | Difference vs Reference |
|---|---|---|---|---|---|
| Coarse | 5,154,755 | 30 | 7 | 0.282 | -7.08% |
| Medium | 14,043,191 | 10 | 10 | 0.300 | -1.12% |
| Fine | 22,979,588 | 5 | 15 | 0.302 | -0.49% |
Flow-Field Assessment
In addition to integrated drag metrics, the simulations produced surface-pressure and velocity-field outputs consistent with external automotive aerodynamics. The solver captured attached boundary-layer development, flow acceleration over the roof and side structures, rear wake formation, pressure recovery, and wheel-wake and underbody interaction regions.
Workflow Integration
The validation workflow was executed through Gradient Dynamics Studio, covering geometry import, automated mesh generation, boundary-condition setup, GPU-native CFD simulation, drag coefficient extraction, and flow-field visualisation. The workflow also supports export to established post-processing tools including ParaView and Tecplot.
Conclusion
The results demonstrate benchmark-quality aerodynamic prediction on complex road-car geometries using a GPU-native CFD workflow designed for high-throughput simulation and engineering optimisation. The solver matched published DrivAerML reference data across multiple geometry variants and showed predictable convergence under mesh refinement.
