Investigation of thermal-hydraulic performance of metal-foam heat sink using machine learning approach

Abstract

Metal foam heat sinks (MFHSs) are often used for electronic cooling because of their high specific heat transfer surface area, fluid mixing capability, and lightweight. Thermal-hydraulic performance of MFHS is known to be related to the morphological parameters of the metal foam (MF), physical dimensions of the heat sink, and coolant flow condition. Moreover, the highly random and complex geometrical nature of MFHS makes full-scale numerical simulation and experimental analysis expensive. In this paper, we have utilized machine learning (ML) methods to predict the thermal-hydraulic performance of MFHS, considering MF morphology, heat sink dimension, and coolant flow condition as input features. Considering local thermal non-equilibrium conditions for MFHS, a database of 1000 data points is generated utilizing a validated high-fidelity computational fluid dynamics/heat transfer (CFD/HT) model. The data sets consist of two non-dimensional descriptors, i.e., Nusselt number (Nu) and friction factor (f), along with the four input features such as MF morphological parameters, i.e., porosity and pore density, heat sink geometrical dimensions, i.e., hydraulic diameter (Dh) and flow condition, i.e., Reynolds number (Re). Five different ML-based regression models, namely k-nearest neighbor (KNN), random forest (RF), extreme gradient boosting (XGBoost), support vector regressor (SVR), and artificial neural networks (ANN), have been developed. Hyper-parameters of each ML model have been systematically optimized, and optimized model's performance have been compared using statistical score metrics and considering the initialization effect. Additionally, in order to demonstrate the robustness of the developed ML models, individual model performance has been evaluated by independent datasets. Results indicate that for testing datasets, all ML algorithms, except KNN, can predict the thermal-hydraulic performance of MFHS reasonably well, with a mean absolute percent error (MAPE) of about 4.59% and a 1.24 standard deviation (SD). However, SVR and ANN outperform other models for independent datasets with a MAPE of less than 3.1%.