Yilei Wei, Xin Yang, Peixin Wu, Guoyou Liu
The presence of voids in the solder layer of an IGBT power semiconductor module can lead to localized hot spots and increased junction temperature during operation, thus posing a threat to the reliability of the power module. Therefore, accurately establishing the mapping relationship between void characteristics and chip temperature is of great significance for reliability evaluation. Based on the complexity of the task and the excellent global optimization ability of the genetic algorithm (GA), the GA-Levenberg–Marquardt (LM)-backpropagation (BP) neural network is used as the basic model framework. Ahierarchical intelligent optimization algorithm is proposed to initialize the hyperparameters. Then, a temperature prediction model is established. The proposed model has an excellent fitting degree, and the determination coefficient of test set is 99.49%. Compared with the other three optimization models, the proposed model has the largest proportion of predicted values in low error ranges. 96.7% of the measured and predicted maximum chip surface temperature ($\text {T}_{\text {m}}$) data are within an error range of [–$\text{2}~^{\circ }$C,$\text{2}~^{\circ }$C] while 97.3% of the measured and predicted average active-area chip surface temperature ($\text{T}_{\text {a}}$) data are within an error range of [–$\text{1}~^{\circ }$C,$\text{1}~^{\circ }$C]. Finally, IGBT modules with different void characteristics are produced by a vacuum reflow soldering system. The applicability of the proposed model is validated by an infrared (IR) camera and thermosensitive electrical parameter (TSEP) method. The results show that the errors of$\text {T}_{\text {m}}$and$\text {T}_{\text {a}}$are within$\text{2}~^{\circ }$C and$\text{3}~^{\circ }$C, respectively, reaching a satisfactory range.