These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Aerodynamic analysis plays a critical role in vehicle design, as it directly impacts fuel efficiency, stability, and overall performance. Traditional computational methods, such as CFD, can provide accurate predictions but are computationally expensive, particularly for complex geometries like those found in cars. Several machine learning (ML) models have been proposed in the literature to significantly reduce computation time while maintaining acceptable accuracy. However, ML models often face limitations in terms of accuracy and scalability and depend on significant mesh downsampling, which can negatively affect prediction accuracy. In this work, we propose a novel ML model architecture, DoMINO (Decomposable Multi-scale Iterative Neural Operator) developed in NVIDIA Modulus to address the various challenges in modeling the external aerodynamics use case. NVIDIA Modulus, which is a state-of-the-art, open-source scientific machine learning platform that enables research, development and deployment of surrogate ML models for a wide range of CAE applications. DoMINO is a point cloud-based ML model that uses local geometric information to predict flow fields on discrete points. It takes the 3-D surface mesh of the geometry in the for of an STL as an input. A 3-D bounding box is constructed around the geometry to represent the computational domain. The arbitrary point cloud representation is transformed into N -D fixed structured representation of resolution m × m × m × f defined on the computational domain. A multi-resolution, iterative approach described in the next section is used to propagate geometry representation into the computational domain and to learn short- and long-range dependencies. Next, we sample a batch of discrete points in the computational domain. When the model is training, these can be points where the solution is known (nodes of the simulation mesh) while during inference they can sampled randomly as the simulation mesh may not be available. For each of the sampled points in a batch, a subregion of size l × l × l is defined around it. A local geometry encoding is calculated in this subregion. The local encoding essentially a subset of the global encoding depending on its position in the computational domain and is calculated using point convolutions. Furthermore, for each of the sampled points in a batch, p nearest neighboring points are sampled in the computational domain to form a computational stencil of p + 1 points. The batch of computational stencils are represented by their local coordinates in the domain. The local geometry encoding is aggregated with the computational stencils to predict the solutions on each of the discrete points in the batch The DoMINO model is trained and evaluated on the DriveAerML dataset [1]. Through our experiments we will demonstrate the scalability, performance and accuracy of our model as well as the scalable and performant end-to-end data, training and testing pipelines that we have developed for accelerated computing. We will also introduce a benchmarking utility developed in NVIDIA Modulus to compare DoMINO on several CAE specific metrics with other model architectures for the external aerodynamics use case. Finally, we will demonstrate the end-to-end, real-time inference and visualization of DoMINO using NVIDIA Omniverse. 1. Ashton, N., Mockett, C., Fuchs, M., Fliessbach, L., Hetmann, H., Knacke, T., ... & Maddix, D. (2024). DrivAerML: High-fidelity computational fluid dynamics dataset for road-car external aerodynamics. arXiv preprint arXiv:2408.11969.