Data-driven prediction for heat transfer and fluid flow through random cylinder arrays at low Reynolds number

Mini- or micro-sized cylinder arrays, such as pin-fin heat sinks, are popularly used to cool overheated miniature devices. Many studies have focused on finding the optimal structural design to enhance cooling performance while maintaining low pressure drop by considering the effect of pitch distance, shape, and size. However, previous studies have primarily focused on heat transfer in homogeneous cylinder arrays with uniform spacing and size. In this study, we introduce a new machine learning technique to predict heat transfer and fluid flow through random cylinder arrays, considering heterogeneous configurations, at low Reynolds number (Re≤100). To generate the training dataset systematically and efficiently, we employed recursive computational fluid dynamics (CFD) simulations without the need to create an excessive number of training domains to account for heterogeneity. A high-order approach was developed to address spatial interdependency, including wake and proximity interference effects between neighboring cylinders. The prediction performance was verified by comparing the predicted results to the CFD results for a wide range of cylinder arrays with random pitch distances and cylinder sizes. Furthermore, our findings demonstrate the developed approach can be effectively applied to the diverse working fluid by considering their physical properties, quantified by the Prandtl number (1≤Pr≤5).