Thermo-hydraulic optimization and rapid temperature field prediction for high heat flux density microchannel heat sinks based on multitask artificial neural networks

High-power electronic devices generate increasingly high and non-uniform heat fluxes, placing stringent demands on microchannel heat sinks to achieve both effective thermal management and low flow resistance. However, conventional CFD-based trial-and-error design is computationally expensive, while existing surrogate models typically focus on scalar thermal–hydraulic performance prediction and have limited capability to provide spatial temperature-field information. To address this issue, this study proposes a three-zone non-uniform staggered pin-fin microchannel heat sink and develops an integrated surrogate-assisted optimization framework that couples a multi-task learning artificial neural network (MTL-ANN), proper orthogonal decomposition (POD), and the non-dominated sorting genetic algorithm II (NSGA-II). Twelve geometric parameters in the inlet, middle, and outlet regions are selected as design variables, and a training dataset is generated using Latin hypercube sampling and conjugate heat-transfer CFD simulations. The MTL-ANN simultaneously predicts pressure drop, maximum heat-source temperature, and temperature uniformity, while the POD model further enables reconstruction of the two-dimensional heat-source temperature field. CFD validation of seven optimized designs shows that the maximum prediction error of the surrogate model is below 5.4%. At a Reynolds number ( R e ) of 402, the All-Balance optimal design achieves a peak performance evaluation criterion (PEC) of 1.34, effectively balancing heat-transfer enhancement and flow-resistance reduction. The first 10 POD modes capture 99.95% of the total temperature fluctuation energy, with an average temperature-field reconstruction error below 1.64%. Compared with conventional CFD, the surrogate model reduces the single prediction time to less than 2 s while simultaneously reconstructing the two-dimensional bottom-wall temperature field, corresponding to an approximately 4500-fold improvement in computational efficiency. The proposed framework provides an efficient surrogate-modeling approach for multi-objective intelligent design and rapid two-dimensional temperature-field evaluation of microchannel heat sinks.