To reduce the computational effort only half of the complex but symmetric geometry of the test facility is simulated. According to the experience gained from the numerical analysis of SUCOS-2D, all structures have to be modeled in detail and a very fine grid is necessary for a good resolution of the thin boundary layers near walls and coolers. The mesh size of the grid changes from 8 mm to 1 mm. Ratios of the mesh sizes between two neighboring cells are less than or equal to 2. The 3d grid consists of 691,000 fluid cells and 68,000 structure cells. A first order upwind scheme is used to compute the convective fluxes of enthalpy and momentum. No turbulence model is used because the flow was laminar in the experiment.

The connecting tubes of the coolers, which are present in the horizontal side area are spatially recorded and modeled even if it is expected that they would have a minor influence on the natural convection than in the calculation of SUCOS-2D. The heat losses to the outside through the lateral walls are neglected. The active heat exchangers can be modeled by a heat exchanger model or by pre-setting a distribution of surface temperature or of heat flux. The calculation of SUCOS-2D showed that it is not necessary to simulate the coolers with the complex heat exchanger model. It is sufficient to give a distribution for the temperature on the surface between fluid and cooler. A linear distribution for the temperature is approximated by a step function prescribing three values for the vertical right coolers. For the horizontal coolers a constant value of temperature is sufficient because the difference of temperatures between inlet and outlet coolant water is less than 1 K. The prescribed values are determined by means of experimental data.

In former simulations for SUCOS-2D it was found that the heated copper plate needs special attention (Kuhn 1996). Even the developing circulation sense in the complete fluid domain is sensitive to the thermal boundary conditions used at the upper surface of the copper plate (Grotzbach et al. 1997). There, the problem of using an artificial Neumann or Dirichlet bound­ary condition was analyzed by calculating the heat conduction in the copper plate. 2d tests showed the surprising result that the copper plate does not ensure a constant heat flux to the fluid, but that it redistributes the heat horizontally in such a strong manner, that the heat flux

into the fluid varies along the plate surface by more than +/-50% of its mean value, Fig. 4. Thus, the thermal conduction in the heater plate is also calculated here. A 3d grid is used for the heated plate; the horizontal grid width distribution corresponds to the one of the fluid re­gion; 5 cells are used in the vertical direction with mesh sizes of 6 mm. The electrical heaters below the copper plate are simulated as a heat flux boundary condition with constant horizon­tal distribution.

The simulation was performed on a CRAY J916 with a memory need of 2.7 Gbytes. The tran­sient calculation was preceded by a steady state calculation to obtain an initial flow and tem­perature field for the transient calculation. The steady state calculation is stopped when an equilibrium in the changes of temperature and in the balance of the heat fluxes is nearly achieved. This happened after 4 h corresponding to 240 h of CPU-time. The transient calcula­tion is performed for a problem time of 227 s with a time step width of 1.0 s. This corresponds to 407 h CPU-time. The system of the pressure equations is solved by the iterative CRESOR method (Borgwaldt 1990), whereas the system of the enthalpy equations is solved by the itera­tive SOR method.