Reproducibility of GPU-based Large Eddy Simulations for mixing in stirred tank reactors

Abstract

CFD simulations are widely used to quantify the mixing performance of stirred tanks for various applications in chemical engineering and biotechnology. Due to advances in GPU computing, these simulations increasingly employ Large Eddy Simulation (LES), which explicitly resolves the dynamics of large-scale turbulence. Although such simulations are fully deterministic and therefore theoretically reproducible, small numerical variations induced by round-off errors, floating-point arithmetic, and differences in the distribution and ordering of operations in parallel computing lead to separation of trajectories i.e., different flow-field evolutions and consequently to significant run-to-run variability in predicted mixing times, even on the same hardware architecture. This work investigates the impact of repeated simulations, in the form of a case study, on the mixing-time distribution observed in a (Formula presented) stirred tank reactor using two commercial CFD packages operating with representative, production-level solver configurations. The analysis does not aim to assess the general performance of numerical method classes, but rather to quantify run-to-run variability under fixed solver settings and to compare the resulting numerical distributions to experimental variability. The results demonstrate that numerical variability is of comparable magnitude to the experimental spread, highlighting the necessity to treat LES-derived metrics as statistical ensembles rather than deterministic values. It is concluded that the reporting of confidence intervals is essential for methodological rigour in LES-based mixing studies.