Using the Group Method of Data Handling Neural Network, and the MOGWO Meta-Heuristic Algorithm to Predict the Thermophysical Properties, Heat Transfer, and Friction Factor of Magnetic Nanofluids in a Heat Sink Under a Magnetic Field

Abstract

The purpose of this study is to use the group method of data handling (GMDH) neural network and the MOGWO meta-heuristic algorithm to predict the thermophysical properties, heat transfer, and friction factor of magnetic nanofluids in a heat sink under a magnetic field. The GMDH neural network and the MOGWO meta-heuristic algorithm are combined in this study. The ANN is first fed the experimental data. To better match the expected results with the experimental data and decrease the error, the meta-heuristic method tweaks the ANN's hyperparameters. By adjusting the number of iterations and associated aspects, which greatly affect the effectiveness of meta-heuristic algorithms, this situation was optimized. To find the best mode, we compare them using two metrics: R and RMSE. It was found that, as the Reynolds number increases, the fluid flow changes from a laminar state to a mixed or mixed-solid state. These changes lead to an increase in convection heat transfer, which increases the Nusselt number. Also, in laminar flows, due to the parallel and regular movement of the layers, there is less resistance to the flow, and as a result, the friction factor decreases. As the volume fraction increases, more collisions occur between solid particles and the pipe walls, which leads to an increase in the friction factor. The optimal prediction for Nu is achieved with 80 wolves and 300 iterations. Additionally, the most accurate FF prediction is attained with 50 wolves and 200 iterations. Finally, this situation may cause the flow pattern to change from a calm state to a turbulent state, which will result in a higher friction factor. On the other hand, by reducing the volume fraction, the amount of collision of solid particles with the walls will be reduced and the flow will be calmer and more stable. This suggests that the algorithms were successful in predicting the behavior of the experimental data.