Drop impact analysis of TSV-based 3D packaging structure by PSO-BP and GA-BP neural networks

Particle swarm algorithm (PSO) and genetic algorithm (GA) were used to optimize the back propagation (BP) artificial neural network for predicting the dynamic responses of the through silicon via (TSV) based three-dimensional packaging structures. A finite element model of the TSV packaging structure with a strain-rate dependent constitutive model for solder joints was created to simulate the drop impact due to a free fall of 0.8 m to the rigid ground to investigate the structural dynamic responses during the whole impact process. The spatial coordinates of the solder joints were used as the input parameters of the hybrid neural network model for the drop impact responses, while the maximum Von Mises stress and PEEQ (plastic strain) values are identified the output parameters. The correlation coefficient (R), the mean absolute percentage error (MAPE) and the training time were used as the measures to validate and compare the proposed PSO-BP and GA-BP neural networks. The results show that both the PSO-BP model and the GA-BP model can achieve high accuracy predictions with strong generalization capability. Apparently, both optimized algorithms outperform the original BP model, but the PSO-BP model is slightly more superior than the GA-BP model. It is also demonstrated that the proposed optimized algorithms efficiently predict the drop impact responses of TSV packaging structures by greatly saving the computational and experimental cost of drop impact tests.