Optimization of Microjet Location Using Surrogate Model Coupled with Particle Swarm Optimization Algorithm

This study aimed to present the design methodology of microjet heat sinks with unequal jet spacing, using a machine learning technique which alleviates hot spots in heat sinks with non-uniform heat flux conditions. Latin hypercube sampling was used to obtain 30 design sample points on which three-dimensional Computational Fluid Dynamics (CFD) solutions were calculated, which were used to train the machine learning model. Radial Basis Neural Network (RBNN) was used as a surrogate model coupled with Particle Swarm Optimization (PSO) to obtain the optimized location of jets. The RBNN provides continuous space for searching the optimum values. At the predicted optimum values from the coupled model, the CFD solution was calculated for comparison. The percentage error for the target function was 0.56%, whereas for the accompanied function it was 1.3%. The coupled algorithm has variable inputs at user discretion, including gaussian spread, number of search particles, and number of iterations. The sensitivity of each variable was obtained. Analysis of Variance (ANOVA) was performed to investigate the effect of the input variable on thermal resistance. ANOVA results revealed that gaussian spread is the dominant variable affecting the thermal resistance.