Towards efficient computational fluid dynamics simulations with physics-informed machine learning

Purpose and goal:

Aims and objectives: (1) develop AI methods for efficient computational fluid dynamics (CFD) simulation, which consider problem geometries and symmetries for the best results; (2) evaluate methods from (1) and compare with traditional CFD methods and more naive ML-based solutions; (3) communicate and build networks with research and industrial actors with interests in technologies of type (1). All objectives were met; e.g., we showed that our method (1) is much faster compared to traditional CFD methods and gives significantly better results compared to simpler AI methods.

Expected results and effects:

We showed that our AI method is much faster compared to traditional CFD methods and provides significantly better results compared to simpler AI methods. The project also resulted in strengthened cooperation with the actors in the reference group (SMHI, WIN Guard). Overall, we see great potential in the developed AI approach when it comes to accelerating CFD, which forms the basis for a large number of technologies and solutions that are central to our modern society and its ability to evolve in a more sustainable direction.

Approach and implementation:

We began with a literature review, followed by simulation of data with relevance to practical applications. Then we implemented a fast and efficient so-called equivariant AI model for CFD simulation. We finished with evaluations of the model and wrote a preprint for publication (under review).

We had ongoing discussions with project partners to gain insight into their problems and to build long-term contacts. Despite some technical challenges, the project has been successful; we have increased our understanding and started a long-term relationship with the project parties.