An experimental study integrated with prediction using deep learning method for active/passive cooling of a modified heat sink

Abstract

In this paper, the effects of various cooling methods for electronic chipsets integrated with a deep learning algorithm are studied. For this purpose, the separated and simultaneous impacts of active/passive cooling methods on the thermal management of a modified heat sink are experimentally investigated. As active methods, two modules of circulating liquid flow and fan are employed. For the circulating liquid flow, distilled water and multiwalled carbon nanotubes (MWCNT) nanofluid, with flow rates in the range of 50–150 ml/min, pass through the heat sink which is attached to a printed circuit board (PCB) where heat fluxes in the range of 20000–40000 W/m2 are generated. For the fan, two modes of 4.5 and 9 W are applied in order to generate air forced convection. Regarding the passive methods, a phase change material (PCM) and nano-phase change material (NPCM) are integrated with the heat sink; paraffin is used as a PCM while MWCNT is added into paraffin to prepare NPCMs with mass fractions of 0.5 % and 1 %. The results show that employing a combination of MWCNT nanofluid flow, air forced convection, and NPCM with 0.5 % mass fraction leads to the best thermal performance where the PCB maximum temperatures of 29.9, 34.4, and 39.4 °C are measured under heat fluxes of 20000, 30000, and 40000 W/m2, respectively. Applying a hybrid configuration where all active and passive cooling methods are utilized, may enhance the mean heat transfer coefficient by a substantial amount (e.g., 42.7 % for 20000 W/m2) compared to the case where only circulating water flow is used. This observation demonstrates the high potential of hybrid methods for the purpose of cooling electronic chipsets. Next, to predict the temperature of the PCB during the operation under different working conditions, various deep learning algorithms are considered. These algorithms include reinforcement learning-group method of data handling (RL-GMDH), multi-layer perceptron (MLP), radial basis function (RBF), support vector regression with the linear kernel (SVRL), and Gaussian kernel (SVRG). The RL-GMDH algorithm is found to result in the best prediction for the PCB temperature (up to 30 s during the operation) which agree well with those of the experiments for various intermittent/continuous operation loads. The predictions of this learning algorithm can open a new approach for the controlling systems to prevent overheating of the electronic chipsets.