Numerical difficulties in the calculation of complex turbulent flows

In this paper the numerical accuracy of a finite-difference technique typical of many used for the calculation of elliptic turbulent flows is discussed. The size of the inevitable truncation terms are investigated both analytically and by direct comparison between predictions and experiment. A classic turbulent shear flow is considered first and it appears that although local errors can be large the general features of the flow are well predicted. Flow down a rearward facing step and over a two-dimensional fence are investigated next. It is clear that inadequate turbulence models are not the sole, or even perhaps the major, cause of discrepancies between prediction and experiment. In particular, it is concluded that numerical errors in regions of difficult geometry (e.g., sharp corners) can be severe so that it seems naive to suppose that the more common numerical methods using 'blanket' grids of the size often used are capable of accurate predictions.