Data-driven Model Discovery for Turbulent Flows (Turbulence Discovery)

Accurate predictions of turbulent flows are of paramount importance for the design and optimization of many engineering systems such as aircraft, wind farms, and heat exchangers. The central issue here is the direct numerical simulation of turbulence is prohibitively costly due to the need for resolving a wide range of scales in space and time. This has left researchers with only one viable option to relax the computational cost: turbulence modeling. However, current turbulence models are built with several simplistic assumptions and, thus, fall short of predicting complex real-world flows. To overcome this long-standing shortcoming, we propose progressive and physics-integrated machine learning (ML) as a novel approach to developing cutting-edge turbulence models. Emphasis is placed on interpretability and generalization, which are crucial to avoid a lack of trust in models by end-users. The approach differs from traditional ML methods by mimicking the progressive development of empirical models based on physical hypotheses, rather than blindly using ML forgetting established physical laws. Our vision is to transform the role of data-driven fluid engineering as a feasible and predictive tool.