Study on Using Noisy Synthetic Data for Neural Networks to Assess Thermo-Mechanical Reliability Parameters of Solder Interconnects

In this study, a feasibility study on using synthesised and augmented data to train and validate an artificial feed forward neural network for the purpose of predicting solder joint stresses due to vibration loads is presented. Data were synthesised by using a full 3D finite element model to extract equivalent elastic strains of Flip Chip solder joints exposed to harmonic vibration with varied amplitude and temperature. The Flip Chip model was varied by means of solder joint size, PCB and chip thickness as well as solder joint population. The synthesised data was augmented by adding Gaussian noise to the input parameters PCB thickness, solder joint diameter, vibration amplitude and test temperature as well as to the calculated strain result to actual noise from manufacturing and measurements. It is shown that training and validation could be successful done with prediction errors less than 5%.