Predictive accurate machine learning models for aerodynamics
| Reference number | |
| Coordinator | Scania CV AB |
| Funding from Vinnova | SEK 824 649 |
| Project duration | December 2024 - October 2025 |
| Status | Completed |
| Venture | FFI ASP - Accelerated Startup Partnership: increased innovation collaboration between startups and need owner |
Important results from the project
The project established a non-parametric geometry workflow for aerodynamic evaluations. A geometry autoencoder compresses detailed truck meshes into a compact latent, which is coupled with operator-based predictors to estimate target fields (cabin; cabin+trailer external surfaces). A pre-alpha engineering platform packages these capabilities for practical use.
Expected long term effects
The results pave the way for faster aerodynamic evaluations at assembly scale using data-efficient AI. By learning directly from non-parametric truck geometries, the approach can reduce the number of full CFD runs needed for early screening and shorten iteration time. The methods are being integrated into a platform designed for engineers, supporting gradual expansion from cabin-level use to larger assemblies as data coverage increases.
Approach and implementation
Team & Roles: From Scania the product owner/technical lead was Thomas Hällqvist. From PredictiveIQ the project was led by Juan F. Betts (CEO) and Fermin Mallor Franco (CTO). Data & Computing setup: CFD data from PowerFLOW simulations was used as the input for the project; with the AI infra hosted on AWS and provided by PredictiveIQ. Methods & models: Neural-operator surrogates of the DeepONet family were used.