Low Cost Recurrent and Asymptotically Unbiased Estimators of Statistical Uncertainty on Averaged Fields for DNS and LES

In the context of fundamental flow studies, experimental databases are expected to provide uncertainty margins on the measured quantities. With the rapid increase in available computational power and the development of high-resolution fluid simulation techniques, Direct Numerical Simulation and Large Eddy Simulation are increasingly used in synergy with experiments to provide a complementary view. Moreover, they can access statistical moments of the flow variables for the development, calibration, and validation of turbulence models. In this context, the quantification of statistical errors is also essential for numerical studies. Reliable estimation of these errors poses two challenges. The first challenge is the very large amount of data: the simulation can provide a large number of quantities of interest (typically about 180 quantities) over the entire domain (typically 100 million to 10 billion of degrees of freedom per equation). Ideally, one would like to quantify the error for each quantity at any point in the flow field. However, storing a long-term sequence of signals from many quantities over the entire domain for a posteriori evaluation is prohibitively expensive. The second challenge is the short time step required to resolve turbulent flows with DNS and LES. As a direct consequence, consecutive samples within the time series are highly correlated. To overcome both challenges, a novel economical coprocessing approach to estimate statistical errors is proposed, based on a recursive formula and the rolling storage of short-time signals.