2022 UK Fluids Conference, Sheffield

In recent years many data assimilation techniques have been introduced for improved Reynolds-averaged Navier-Stokes (RANS) turbulence modelling. While some rely on huge datasets of high-fidelity data, such as those available from direct numerical simulations (DNS), other frameworks have been shown to achieve considerable improvements using limited, experimentally measurable, data. One such framework is field inversion: an approach that involves perturbations of the turbulence model transport equation through a spatial field and the iterative optimisation of this field such that the error between model prediction and data is minimised. Results reported in literature, and ours, have been primarily dictated by the availability of data for particular test cases, usually generated for benchmarking and validation (i.e. not specifically tailored for data-driven turbulence models). It has been variously observed that the impact of field inversion in terms of error reduction in all physical quantities of interest are highly sensitive to the type of data used. However, there has been no quantitative study of the relative importance of physical quantities for data-driven turbulence modelling. In this work we will attempt to tackle this. We will use limited data from rich DNS datasets to emulate experimental scenarios (e.g. pressure taps, hot-wire anemometers, or PIV measurements), and perform field inversion. We will then compare the data-driven turbulence model predictions against the rich DNS data in terms of physical quantities of interest (e.g. surface pressure, skin friction, velocity field), and potentially, turbulent quantities (e.g. turbulent kinetic energy, Reynolds shear stress etc.). This will allow us to use error metrics to analyse the relative importance of different quantities, and will be the first step towards creating a list of ingredients for the experimental fluid dynamicist, and bridge the gap between experimental and computational approaches.