PaddleScience Presents Geometry-Informed Neural Operator Model for Aerodynamic Drag Prediction at 2024 CAE Aerodynamics Branch Academic Annual Meeting

From September 5 to 6, 2024, the Academic Annual Meeting of the Aerodynamics Branch of the Chinese Society of Automotive Engineering was successfully held in Chongqing, co-organized by Chongqing Changan Automobile Co., Ltd., China Automotive Engineering Research Institute Co., Ltd., and the National Key Laboratory of Intelligent Automotive Safety Technology. The meeting is a premier academic platform integrating industry experts and discussing cutting‑edge topics in automotive aerodynamics.

The event attracted over 300 technical experts and scholars from more than 100 domestic and overseas enterprises, research institutions and universities, covering fields such as whole‑vehicle and component manufacturing, CFD software development, AI technology application, and wind‑tunnel construction, with a special focus on the application of artificial intelligence in the automotive industry.

PaddleScience showcased its research achievement entitled “Parameterized Car Geometry Drag Prediction Model Based on Geometric Information Neural Operator” at the conference. The paper was reviewed on‑site and received the Outstanding Paper Award.

The model is built on the SAE aerodynamic standard and a self‑created parameterized 3D CFD dataset (SAE‑Variable) generated with OpenFOAM. It employs the Geometric Information Neural Operator (GINO) to predict surface pressure and shear stress, forming a surrogate model with 2.3 billion parameters. On test data the average relative error in drag‑coefficient prediction is 2.1 %, while the inference speed is three orders of magnitude faster than a full OpenFOAM simulation, enabling rapid aerodynamic shape optimization.

Following the presentation, the PaddleScience team plans to extend the GINO‑based approach to aerodynamic optimization for various vehicle types, establishing a full‑process workflow that leverages AI to improve design efficiency.

PaddleScience is a scientific computing toolkit based on the deep‑learning framework PaddlePaddle. It utilizes deep neural networks and PaddlePaddle’s automatic (high‑order) differentiation to solve problems in physics, chemistry, meteorology, etc., supporting physics‑driven, data‑driven, and physics‑informed methodologies, and provides basic APIs and comprehensive documentation for users and secondary development.

The toolkit offers rich classic case studies (e.g., 2D/3D cylinder flow, vortex‑induced vibration, weather forecasting, pollutant dispersion), extensive API updates (custom PDEs, various boundary conditions, 2D/3D primitive geometry definitions), framework innovations in automatic differentiation, compilation, execution, and distributed computing, and broad model support including CNN, U‑Net, and Transformer architectures.

To foster a closed loop of industry, academia, and research, PaddlePaddle has partnered with enterprises, universities and research institutes to launch the open‑source AI for Science initiative, inviting researchers to visit the PaddleScience official site and join the AI for Science co‑creation plan.