Minisymposium

In the era of LLM models, one gets notoriously confronted with the question of where we stand with applicability of large-scale deep learning models within scientific or engineering domains. The discussion starts by reiterating on recent triumphs in weather and climate modeling, making connections to computer vision, physics-informed learning and neural operators. Secondly, we discuss challenges and conceptual barriers which need to be overcome for the next wave of disruption in science and engineering. We showcase recent breakthroughs in multi-physics modeling, computational fluid dynamics, and related fields.

In this talk, we will focus on discussing recent work to assess the capability of AI surrogate models in the prediction of automotive aerodynamics. In particular, the talk will focus on both dataset generation (specifically efforts behind generating the AhmedML, WindsorML and DrivAerML open-source datasets) as well as results from both GNN and neural operator approaches for surface, volume and integral quantities. The focus will be on bringing to light successes but also failures and challenges in current methods that can hopefully motivate further development by the community.

In areas for which vast amounts of data are available Machine learning and artificial intelligence techniques had a tremendous success, especially when mathematical models are lacking. Instead, engineering tools in general and computational fluid dynamics tools in particular rely on first-order principals that directly enable to describe and investigate system behavior. However, such tools are far from perfect and suffer several short-comings, e.g. computational bottlenecks once a massive amount of simulations is required or the problem of deriving accurate turbulence models to describe small scale turbulent behavior. Machine learning techniques are generally regarded as a possibility to enhance and complement first-order based numerical simulation tools to circumvent these shortcomings. Following this ambition, the Center for Computer Applications in AeroSpace Science and Engineering department of the German Aerospace Center investigates scientific machine learning techniques in close connection to established numerical simulation tools as well as industrial needs. This presentation will provide an insight into previous and current activities within the department covering topics from purely data-driven approaches to the incorporation of physical knowledge into models specifically highlighting challenges that arise when looking at large-scale industrial configuration and integration into existing, traditional workflows.

In this panel session, we will bring together the speakers to discuss the development of ‘Foundational Models’ a key topic in the field of Machine Learning for Computational Fluid Dynamics. In particular, we will discuss the technical challenges surrounding data generation, generalizability, ML architecture choice as well as broader commercial and legal concerns. We will invite the audience to share in this panel discussion and have a debate on aspects surrounding accuracy, computational cost, commercialization and overall feasibility.