Reproducibility of Large Eddy Simulations for mixing in stirred tank reactors

CFD simulations are widely used to quantify mixing performance of stirred tanks, for various applications in chemical engineering and biotechnology. Due to advances in GPU computing, more and more often these simulations make use of Large Eddy Simulations (LES), which explicitly simulate the dynamics of large-scale turbulence. Although these simulations are fully deterministic and hence theoretically reproducible, small numerical variations induced by round-off errors combined with differences in distribution and order of operations in parallel computing lead to separation of trajectories, i.e. different flowfield evolutions and different mixing times between repeat simulations, even on the same architecture. We investigate the impact of repeat simulations on the mixing time distribution observed in a $30 \liter$ stirred vessel with two commercial CFD packages, and compare to experimental variability. While the distribution between simulations and experiments is in very good agreement, we do conclude confidence intervals should be reported for CFD simulations of mixing.