Mitigating weaknesses of density-based thermo-fluid topology optimization

Abstract

Power electronic systems are reaching higher efficiencies as their technology advances, which often results in components of smaller size with higher power densities. Cooling these components becomes increasingly challenging as high power densities require cooling with large heat fluxes. Topology optimization (TO) of thermo-fluids can be used to find cooling interface geometries which achieve high heat transfer with realistic pumping power. However, current methods for thermo-fluid TO show several issues. Firstly, the fluid models used for thermo-fluid TO show weaknesses that cause deviating behavior compared to conventional fluid models, which can result in large under- or overestimation of heat transfer especially when applied to turbulent flow. Secondly, although this deviating behavior is known to result in inferior modeling accuracy, the achievable accuracy of thermo-fluid TO has never been quantified. Lastly, TO currently requires many modeling parameters to be specified manually. Since these parameters largely affect the accuracy of the thermo-fluid solver, tedious parameter tuning is part of the TO design process. This thesis firstly presents a framework which allows quantitative analysis of the modeling accuracy achievable with density-based thermo-fluid models in 2D. The framework reveals several effects causing errors in turbulent flow, as well as a predictability of density-based boundary layer flows. Secondly, a method is tested which minimizes errors of the density-based thermo-fluid models by meta-optimizing the modeling parameters. Applied to a test-case with turbulent flow, this method achieves up to 27% reduction of the modeling error compared to a parameter sweep. When applied to a laminar TO, it achieves similar accuracy as a manually tuned TO without needing any tuning. A second mitigation method which adjusts the thermal conductivity of porous solid material to compensate for erroneous convective heat transfer is infeasible, as it provides less accurate results than the first method. Lastly, a post-processing method which uses the meta-optimized data is tested and found to provide better accuracy than a conventional post-processing method.