TC-thermal: A novel hybrid transformer-CNN architecture enhancing thermal flow reconstruction at fluid-solid interface for micro-channel heat sink

Acquiring instantaneous temperature and velocity distributions is critical for the thermal management of microchannel heat sinks in electronic devices. Although traditional machine learning data-driven approaches are capable of rapidly predicting physical fields, they often fail to precisely characterize local features in regions of high temperature gradient, particularly at the fluid-solid interface where hotspots predominately occur. To address such limitation in prediction performance at the interface, this work proposes a novel data-driven model named TC-Thermal, which integrates a convolutional neural network (CNN) encoder leveraging a self-attention mechanism decoder to predict the temperature and velocity distributions of a microchannel cooling system. By means of TC-Thermal, both temperature and velocity field are well reconstructed only with heat fluxes, inlet conditions and flow rates provided. The mean absolute percentage errors (MAPE) are as low as 0.02 % and 2.3 % for temperature and velocity prediction, respectively, compared with the numerical results. Particularly, the mean absolute error (MAE) of temperature prediction at the fluid-solid interface is 0.17 K, representing 22.27 % and 58.03 % improvements over the standalone Transformer model and the CNN model, respectively. The results demonstrated our proposed TC-Thermal architecture in capable of both capturing global and local thermal properties, which contributes to the application of data-driven method for predicting thermal dynamics with high temperature gradient.