Optimization of nanofluid heat transfer in a microchannel heat sink with multiple synthetic jets based on CFD-DPM and MLA

Abstract

This study uses a combination of computational fluid dynamics consisting of the discrete phase model (CFD-DPM) coupled with support vector regression-particle swarm optimization (SVR-PSO) techniques to maximize the cooling performance of Al₂O₃water nanofluids in a microchannel heat sink (MCHS) with multiple synthetic jets (SJs). First, a CFD parametric study is carried out to understand the effect of influential parameters, including orifice spacing, particle volume fraction, orifice height, diaphragm length, phase actuation of jets, frequency and amplitude of oscillating diaphragms. A hybrid SVR-PSO algorithm is then adopted to predict the optimum values of the influential parameters for the maximum homogeneous heat transfer. The results predicted by the machine learning algorithm (MLA) are also compared to CFD findings. 0.15% and 3.5% deviations are found between predicted and actual values for the minimum average temperature and temperature uniformity, respectively. The parametric study reveals that heat transfer increases when the two jets are placed apart from each other. Based on the parametric study, the average temperature drops by 4.3 K as the membrane length increases from 0.95 mm to 1.96 mm. The highest heat transfer is obtained at a particle volume fraction of 5% in both jet arrangements. It is found that increasing the amplitude and frequency of the membranes results in better cooling performance. The results also confirm that larger orifice heights allow the creation of longer and stronger flows in the orifices that propagate the furthest in the microchannel. Overall, in-phase jet configurations show more uniform and lower temperatures (e.g., better heat transfer) at higher particle concentrations compared to the 180° out-of-phase jet arrangements.

Introduction

Future generations of microchips will be designed depending on the size and efficiency of devices that require control of the allowable operating temperature to avoid thermal failure in microchips. Jet impingement as a powerful active method is used for heat removal from a hot surface [1]. In this technique, coolants need to be constantly pressurized by an external supplier. MCHSs can dissipate heat from microprocessors. Laminar flows in MCHSs create low pressure drops, though the heat transfer rate is not large enough. To compensate for this deficiency, Timchenko et al. [2] proposed integrating SJs to MCHSs to disrupt the laminar flow, to generate better mixing, that enhances the heat transfer. Periodic flows produced by a diaphragm oscillating in a cavity generate fluid suction and ejection through an orifice across the flow boundary. Lee et al. [3] studied flow and heat transfer in a microchannel equipped an SJ. The SJ demonstrated remarkable heat transfer enhancement over a silicon substrate without causing associated changes in other operating parameters such as pressure drops. A quasi-steady condition was also observed for the single SJ. Lee et al. [4,5] further developed multiple SJs in a microchannel to produce rigorous flows and promote the uniformity of local temperature distributions. They found that the heat transfer rate was maximized by 180° out-of-phase configurations at a frequency of 560 Hz. This is consistent with recent studies [6,7] confirming that the flow structure is highly influenced by the oscillation of the SJs. A study conducted by Gil et al. [8] also showed that generating strong vortices led to higher heat transfer in SJs.

To further enhance the heat transfer, a mixture of solid nanoparticles and base fluids was introduced as nanofluids by Choi et al. [9]. Nanofluids provided greater heat transfer compared to base fluids due to their higher thermal conductivity [10]. It was reported that the absorption of nanofluid heat transfer in MCHSs was increased when compared with lower flow rates [11,12]. Murshed et al. [13] affirmed the significant role of nanofluids in the revolution of cooling technologies in the future of the electronics industry. It was shown that smaller nanoparticles had better heat transfer performance compared to larger ones [14,15].

Nanofluid heat transfer has been numerically studied by different models, which can be classified into two main categories: single-phase models (SPMs) and multi-phase models (MPMs). The SPMs are very simple and fast in convergence because, unlike the MPMs, they assume one effective phase instead of considering two solid and liquid phases [16]. Alternatively, tracking dispersed particles in base fluids and exploring different forces acting on particles can be achieved using a discrete phase model (DPM), which is an MPM that uses an Eulerian framework for the base fluid and a Lagrangian approach for particles [17].

The DPM approach was employed in a numerical study of a laminar flow regime in a microchannel [18]. The authors found that the non-uniform distribution of particles near walls affected the fluid flow and heat transfer. They also reported that the DPM model could improve predicting heat transfer compared to the SPM. Ahmadi et al. [19] validated the DPM model through their experimental findings. They introduced the motion and position of nanoparticles as the two most important factors in heat transfer enhancement. Sharaf et al. [20] revealed that the shear rate and Thermophoretic forces had dominant effects on the motion and position of particles near walls and consequently on heat transfer. Mohammadpour et al. [17] applied the DPM model to nanofluids in an MCHS with an SJ. They found that heat transfer was enhanced by decreasing the particle size, increasing particle concentrations, and raising membrane amplitudes and frequencies.

Despite the success of DPM for nanofluids, this method is computationally expensive, particularly for comprehensive parametric studies. The integration of CFD models with machine learning algorithms (MLAs) allows the effects of various parameters that contribute to the optimum solution to be studied simultaneously while significantly reducing the computational cost. In recent years, applications of meta-heuristic algorithms have become widespread in heat transfer studies. Yildizeli et al. [21] considered pumping power and the Nusselt number as two objective functions to evaluate flow and heat transfer in an MCHS. They used a genetic algorithm (GA) to optimize values for the inlet Reynolds number and MCHS geometric parameter. Naphon et al. [22] combined an MCHS, nanofluids, and jet impingement as three effective cooling technologies. They adopted an artificial neural network (ANN) algorithm to optimize the heat transfer and pressure drop. Keykhah et al. [23] worked on the heat transfer and fluid flow of homogeneous nanofluids in a tube. They optimized results through two objective functions (i.e., the Nusselt number and friction coefficient) using a multi-objective particle swarm optimization (MOPSO) algorithm. Siavashi et al. [24] optimized the predicted heat transfer and pressure drop of a fluid flow in a pipe filled with porous media using the PSO algorithm. They found that using the alumina-water nanofluid with a volume fraction of 5% in optimized conditions resulted in better performance evaluation criteria (PEC) values.

This study aims to obtain the highest heat transfer rate in an MCHS through the optimization of significant parameters contributing to the integration of SJs and nanofluids. Here, a CFD- MLA model, for the first time, is developed for such nanofluid configuration. For this purpose, a parametric CFD study is first conducted through the Eulerian-Lagrangian approach (DPM) to identify parameters influencing the heat transfer rate and its uniformity. The CFD dataset is then processed for training the MLA model. In the MLA step, the SVR is adopted to predict heat transfer in the MCHS. In the next step, a PSO algorithm is adopted to find optimum values of the influential parameters to maximize the cooling performance. Eventually, the attained values from the MLA are validated with CFD-DPM results to evaluate the accuracy of the prediction of the SVR-PSO model.