Architected Materials Informatics for Designing Cellular-Structured Heat Sinks Fabricated by Additive Manufacturing

With the increasing integration of electronic devices, there is a need for higher-performance heat sinks to dissipate heat efficiently. Heat sinks usually have fin or pin shapes and enhance the heat transfer at the solid/fluid interface by expanding the surface area. The conventional manufacturing techniques for heat sinks including extrusion, forging, and machining have limited their manufacturable shapes. Recent developments in additive manufacturing (AM) have expanded the freedom of manufacturable shapes and enabled the fabrication of complex-shaped architected materials including lattice or cellular structures. The cellular structures have a larger surface area than fin or pin structures, and more efficient heat transfer will be expected. When heat sinks are used under forced convection, a fluid is supplied to heat sinks by using a fan or rotating heat sinks. The pressure loss needs to be suppressed to supply a fluid with a high velocity efficiently. It is known that high-heat transfer rate and low-pressure loss have a trade-off relationship under a constant fluid velocity because a large surface area increases the resistance of fluid flow. Therefore, it is necessary to develop a methodology for optimizing cellular-structured heat sinks with a good balance of high-heat transfer rate and low-pressure loss according to the high manufacturing freedom of AM processes.<br/>Computational fluid dynamics (CFD) simulations are usually used for evaluating the heat transfer rate and pressure loss of cellular-structured heat sinks. However, the CFD has a high computational cost, and a huge amount of time will be required to repeat the simulations until the heat sink structure is optimized. To reduce the calculation cost, machine-learning surrogate models have been developed and inversely analyzed by optimization algorithms. However, to input structural information into a surrogate model, the structure needs to be converted into numerical data. Therefore, only the dimensions of cellular structures (size of basic structure, thickness of struts, space between adjacent basic structures) have been optimized while the basic structure has been fixed.<br/>In this study, an attempt was made to optimize the basic structures by combining a machine-learning surrogate model and Voronoi tessellation. The Voronoi tessellation divides planes or spaces by arranging seed points and drawing perpendicular bisecting lines or planes between adjacent seed points. When the edges of the perpendicular bisecting planes drawn on the spaces are replaced by solid struts, various cellular structures can be designed by changing the number and coordinates of seed points. In addition, the Voronoi tessellation is compatible with machine learning because the number and coordinates of seed points (numerical data) are uniquely linked to the cellular structures.<br/>800 cellular structures are designed by randomly changing the coordinates of 9 seed points and the thickness of solid struts. The heat transfer rate and pressure loss of the structures were evaluated by CFD simulations under forced convection with a fluid velocity of 1 m/s. The data set of the coordinates of seed points, the thickness of solid struts, heat transfer rate, and pressure loss was trained to a neural network model. The neural network was inversely analyzed by a genetic algorithm (NSGA-II) to obtain a Pareto front of high-heat transfer rate and low-pressure loss. The properties of an optimized structure and a counterpart with a lower heat transfer rate and almost the same pressure loss compared to the optimized structure were evaluated by the CFD simulations and experiments using small-sized wind channel equipment. It was validated in the CFD simulations and experiments that the optimized structure exhibited a higher heat transfer rate and almost the same pressure loss compared to the counterpart. This study constructed a base of a novel methodology for designing cellular-structured heat sinks.