Pmrt Enables Fast, 3D High-Resolution Aerodynamic Prediction, Achieving 7x Cost Reduction with 0.975 Accuracy

Aerodynamic optimisation remains a crucial, yet computationally expensive, challenge in automotive design, demanding close collaboration between engineers and stylists. Sam Jacob, Markus Mrosek, and Carsten Othmer, alongside colleagues at Volkswagen AG and Friedrich-Alexander-Universität Erlangen-Nürnberg, now present a new training methodology, Progressive Multi-Resolution Training (PMRT), that dramatically accelerates the prediction of high-resolution aerodynamic performance. This innovative approach enables the training of a single model to accurately predict drag and detailed velocity fields, achieving results comparable to existing methods but at a significantly reduced cost, a seven-fold reduction in training expense on a single GPU. Crucially, PMRT also demonstrates adaptability, successfully training on data from multiple simulation solvers, including real-world data, and achieving a drag prediction accuracy of 0. 975 on a standard benchmark dataset.

Neural Networks Predict Aerodynamic Drag and Flow

This research details a novel approach to aerodynamic shape optimisation using neural networks, specifically focusing on predicting drag coefficient and detailed velocity fields. Scientists developed a system capable of accurately predicting aerodynamic performance, addressing a significant challenge in automotive design. Models were trained and tested on diverse datasets, including realistic car geometries, simplified shapes, and complex real-world data, demonstrating the versatility of the approach. Careful consideration of the network architecture, including the use of multiple decoders and skip connections, further enhanced predictive accuracy. Recognising that traditional Computational Fluid Dynamics (CFD) simulations are time-consuming and limit design exploration, the team aimed to create a system capable of accurate, real-time aerodynamic predictions. The study pioneered a training schedule that progressively increases resolution, starting with low-resolution data and gradually shifting towards high-resolution fields. This approach utilises a U-Net architecture, a type of convolutional neural network, and trains it by sampling batches from three distinct resolutions based on dynamically changing probabilities.

During training, the probability weighting shifts, progressively prioritising higher resolutions to refine the model’s predictive capabilities. Experiments demonstrate that PMRT achieves this training in just 24 hours on a single NVIDIA H100 GPU, representing a substantial reduction in computational cost compared to training a model solely at high resolution, while maintaining comparable accuracy. Furthermore, the team demonstrated the versatility of PMRT by successfully training a single model across five different datasets, by conditioning the model on simulation parameters. The work addresses a key bottleneck in automotive engineering, where accurate aerodynamic simulations are traditionally slow and expensive. The team achieved the ability to train a U-Net model to predict drag coefficient and velocity fields in just 24 hours on a single H100 GPU, a substantial reduction in cost compared to high-resolution-only methods, while maintaining comparable accuracy. The PMRT approach involves sampling data from three different resolutions during training, starting with lower resolutions and progressively shifting towards higher resolutions.

This strategy allows the model to learn efficiently, prioritising broad features before refining details. Experiments demonstrate that a PMRT model achieves a group drag coefficient Mean Absolute Error of 2. 8 drag counts in 5 hours, outperforming all baseline models. Further optimisation revealed that a PMRT model with specific training parameters lowers error and trains three times faster. When predicting both drag coefficient and velocity fields, the team achieved a group drag coefficient Mean Absolute Error of 2.

5 drag counts and a group relative L2 error of 2. 4%. The team successfully trained a U-Net to predict both drag coefficient and detailed velocity fields, achieving results comparable to existing methods but with up to a seven-fold reduction in training time on a single GPU. This advancement addresses a key bottleneck in aerodynamic optimisation, enabling faster and more efficient design iterations. Notably, the researchers demonstrated the versatility of PMRT by training a single model across datasets generated by different computational solvers, including real-world data.

This capability streamlines the development process and enhances the reliability of predictions across varied simulation environments. The models achieved a correlation coefficient of 0. 975 on the DrivAerML dataset, matching the performance of established benchmarks at a substantially lower computational cost.