Multi-objective reduced-order design optimization of single-phase liquid coolers for electronics

Thermal management using liquid cooled cold plates is a key enabler in modern electronic systems such as power converters and multi-chip arrays. Traditional cold plate designs are limited in their ability to handle the heterogeneous heat profiles in ever-larger modern multi-chip arrays requiring design optimization. This work demonstrates a design optimization method for liquid-cooled cold plates subjected to discrete heat dissipation profiles. We used reduced order performance modeling enabled by splitting the heat sink geometry onto a coarse grid with up to 50 elements each assigned with a simplified flow block such as a straight, elbow, or tee section. Correlations for thermal-fluidic performance metrics were used to obtain the pressure drop and surface temperature. Multi-objective design optimization focused on minimizing the pressure drop and temperature was broken down into layout and geometry optimization sub-steps. Layout optimization found the optimal flow path from inlet to outlet using a novel optimal path search method combining pathfinding algorithms with search on a random population. Geometry optimization solved for the dimensions of the coarse grid elements and the corresponding flow blocks using gradient based optimizers. We experimentally tested an aluminum water-cooled heat sink design generated by the reduced order optimization demonstrating a thermal resistance of 0.065 K/ W. We compared the performance of our optimal designs with conventional and topology optimized cold plate designs to benchmark performance, computational cost, and computational time. Topology optimized designs showed a 22% reduction in average heated surface temperature for similar pressure drops. However, the computation time required to generate the topology optimized design was more than one order of magnitude higher than the reduced order method, requiring over 3,600 s to generate the optimal design. Our reduced order optimization in comparison provides optimal designs at computational timescales of 60 s.