Simulations of computational fluid dynamics are widely used to quantify the mixing performance of stirred tanks for various applications in chemical engineering and biotechnology. Thanks to advances in computing on graphics processors, these simulations increasingly employ large eddy simulation (LES), which explicitly resolves the dynamics of turbulence on the large scales.
Although such simulations are fully deterministic and therefore reproducible in theory, small numerical variations induced by rounding errors, by floating point arithmetic and by differences in the distribution and the ordering of operations in parallel computing lead to a separation of the trajectories. The flow field then evolves differently from one run to the next, and the predicted mixing times vary significantly, even on the same hardware architecture.
The work investigates, in the form of a case study, how repeated simulations affect the distribution of mixing times observed in a stirred tank reactor of 30 litres, using two commercial packages that operate with representative solver configurations at production level. The analysis does not aim to assess the general performance of classes of numerical methods, but rather to quantify the variability from run to run at fixed solver settings and to compare the resulting numerical distributions with the experimental variability.
The results demonstrate that the numerical variability is of comparable magnitude to the experimental spread. Metrics derived from large eddy simulations should therefore be treated as statistical ensembles rather than as deterministic values, and the reporting of confidence intervals is essential for methodological rigour in mixing studies based on this method.
The study is a collaboration between the Institute of Multiphase Flows at Hamburg University of Technology and the Department of Biotechnology at Delft University of Technology, with funding from the Dutch Research Council and the German Research Foundation.
Ryan Rautenbach, Héctor Maldonado de León, Pieter Brorens, Michael Schlüter, Cees Haringa (2026). Reproducibility of GPU-based Large Eddy Simulations for mixing in stirred tank reactors. Computers & Chemical Engineering 210, 109615.
