Pin-fins have been widely used in cooling channels to enhance internal heat transfer. For over thirty years, the literature has been using regular pin-fin shapes or identical pin-fins arrays. However, it was expected that an efficient pin-fin channel should have irregular pin-fin shapes and localized changing shapes along the streamwise direction. These new degrees of freedom for pin-fins were not well explored due to the lack of data processing method in the past. With the aid of the advanced deep learning techniques arising in these years, this study proposed a new optimization approach for pin-fins using the pix2pix networks and the Genetic Algorithms. A simulation dataset with 300 random spline pin-fin shapes was generated using Computational Fluid Dynamics. Two surrogated models were trained and tested to predict the temperature distributions on the external surfaces and pressure distributions in the middle section of the channel. Five optimized geometries were generated using different combinations of cooling objectives and pressure objectives. Based on the comprehensive results proliferated by the machine leaning methods, detailed sensitive analysis and response analysis were conducted for each input parameter. The optimized results indicated several general suggestions for pin-fin designs under different objectives: (a) square pin-fins could fit better with the cooling requirement and pressure constraints for the most upstream regions, (b) enlarging the opening area of middle stream pin-fins could elevate the uniformity of temperature, (c) streamlined pin-fins helped reduce the pressure drop. This effort was expected to provide a reference to explore cooling channel configurations geometries within a larger degree of freedom.

中文翻译：

---

使用深度学习方法寻找适合高温应用的不规则针翅形状


针翅片已广泛用于冷却通道以增强内部传热。三十多年来，文献一直使用规则的针鳍形状或相同的针鳍阵列。然而，预期有效的针翅通道应具有不规则的针翅形状和沿流向的局部变化形状。由于过去缺乏数据处理方法，这些新的针鳍自由度并未得到很好的探索。借助近年来出现的先进深度学习技术，本研究提出了一种使用 pix2pix 网络和遗传算法的针鳍优化方法。使用计算流体动力学生成具有 300 个随机样条针翅形状的模拟数据集。训练和测试了两个代理模型来预测通道外表面的温度分布和通道中部的压力分布。使用冷却目标和压力目标的不同组合生成了五种优化的几何形状。基于机器学习方法得出的综合结果，对每个输入参数进行了详细的敏感分析和响应分析。优化结果对不同目标下的针鳍设计提出了几点一般性建议：（a）方形针鳍可以更好地适应最上游区域的冷却要求和压力约束，（b）增大中流针鳍的开口面积-翅片可以提高温度的均匀性，(c)流线型的针翅片有助于减少压降。这项工作有望为探索更大自由度内的冷却通道配置几何形状提供参考。