Usability Guide#

Use this documentation for technical and usability details of the DoMINO-Automotive-Aero NIM. The users are recommended to follow these guidelines to maximize the accuracy, performance, and generality of the model.

STL configuration#

  • The vehicle STL is supposed to be a single part. For example, separate STLs for body, wheels, spoilers, etc., need to be combined into one before use with the NIM. PhysicsNeMo utilities can be used to combine STLs.

  • STL should be watertight.

  • The STL units are in meters.

  • The STL needs to be oriented such that the axial flow direction is from -x to +x, vertical flow direction from -z to z and side flow direction from -y to +y.

  • The front wheel center of the car is kept at x=0. The wheels should touch the ground plane located at z=-0.318469 m. Finally, ensure that the vehicle’s geometric center lies on the xz plane; i.e., the vehicle is approximately symmetrical about the xz plane. The STL needs to be rotated and/or translated to meet these requirements.

  • The vehicles STLs need to be under 6.3 meters in length, 2.2 meters in height and 2.5 meters in width for the best model performance.

  • The accuracy of the model will depend on the STL resolution, higher resolution corresponds to higher accuracy. Recommended STL resolution > 300k points.

Velocity inlet#

  • Velocity inlet is specified in the flow direction (along the x-axis).

  • Cross flow is not supported.

  • Velocity inlet magnitude should range from 20 m/s to 50 m/s for maximum accuracy.

Stencil size#

  • Stencil size is a parameter to control accuracy of the predictions by modifying the number of neighborhood points sampled around the point where the solution fields are evaluated.

  • Large stencil sizes correspond to improved accuracy, but higher computational costs.

Surface and volume field evaluation#

  • Volume fields are evaluated on the sampling points. These points are uniformly sampled in the computational domain and the fidelity of the results increases as the larger number of points increases.

  • Surface fields are directly evaluated on the STL. Higher resolution STL will correspond to higher accuracy of surface fields and corresponding drag and lift force.

Confidence intervals#

This table aims to provide users with guidance on the model performance across different classes of cars. In the table, we outline the model error metrics between predicted drag force and calculated drag force from OpenFOAM across the different classes of cars. Each test sample represents a unique combination of geometric variation and inflow speed.

Vehicle Class (# samples)

% test samples with error < 5 %

% test samples with error < 15 %

Reliability Score (higher is better)

Sedan (496)

52.6

87.3

220.7

SUV (92)

40.2

85.9

235.9

Hatchback (91)

36.3

92.3

131.3

Pickup (92)

45.7

90.2

140.2

Van (63)

50.8

96.8

138.6

Sports cars (51)

41.2

94.1

97.7

The drag force errors are similar across all the car classes. We provide a Reliability Score metric that combines the percentage errors with the number of topologically unique designs in the dataset. Based on this metric, the model exhibits a higher confidence interval on Sedans and SUVs.

Similarly, the table below outlines the model performance across different speed ranges.

Speed range (# samples)

% test samples with error < 5 %

% test samples with error < 15 %

20 - 29.99 m/s (236)

44.1

89.0

30 - 39.99 m/s (346)

50.9

89.9

40 - 49.99 m/s (294)

49.3

88.1

50 - 59.99 m/s (9)

11.1

88.9

The drag force errors are similar across the various speed intervals, with the speeds between 20-50 m/s showing higher confidence. This is due to lower errors and more testing samples in the 20-50 m/s range.

On this page