Enhancing performance of nanofluid mini-channel heat sinks through machine learning and multi-objective optimization of operating parameters

•The study proposes using CFD and machine learning to optimize nanofluid mini-channel heat sinks.•Numerical models of the mini-channel were generated using the mixture model to develop the dataset for machine learning models.•Support vector regression (SVR), gaussian process regression (GPR), and random forest (RF) models were used to establish the mapping relationships between the parameters of the nanofluid mini-channel.•The GPR model was found to be the most suitable, with high R2 values for both the pressure drop and average temperature of the heating wall.•Using the NSGA-II multi-objective optimization algorithm, the study determined the optimal value of the volume fraction (φ) for different operating conditions, with values of around 3% at low reynolds numbers and around 2% at high reynolds numbers. The improvement in the performance of power systems in new energy vehicles has posed new demands for the performance of thermal management systems, leading to an increased interest in the application of nanofluid mini-channel heat sinks. Despite their potential, studying nanofluids is challenging due to the complexity of their preparation. To mitigate the computational and optimization costs, this study proposed a combination of Computational Fluid Dynamics (CFD) and machine learning in a multi-objective optimization algorithm for optimizing the operating parameters of nanofluid mini-channel heat sinks. Firstly, a numerical model of the nanofluid mini-channel was developed using the Mixture model to generate the dataset for machine learning models. Secondly, SVR, GPR, and RF models were utilized to establish the mapping relationships between the parameters of the nanofluid mini-channel, including the inputs of the inlet Reynolds number (Re), volume fraction (φ), and heat flow density (q), and the outputs of the pressure drop (△P) and the average temperature of the heating wall (Tave). The results indicated that the GPR model was the most suitable, with R2 values of 0.9939 and 0.9985 for Tave and △P, respectively. By employing the NSGA-II multi-objective optimization algorithm, the optimal value of φ was determined for different operating conditions, with values of around 3% at low Re and around 2% at high Re.