Abstract
In recent years, due to the consumer market become larger and larger, the electronic equipment improves every day, we pay more and more attention to the reliability of electronic packaging. Electronic packaging is a complex structure that protects IC devices and plays an important role in the semiconductor industry. Our research aims to evaluate the reliability of electronic packaging, and Thermal Cycling Testing (TCT) is one of the important tests to ensure packaging reliability. TCT is a good way to test electronic packaging reliability, but it takes a lot of time and cost to do this experiment. The time required to perform a TCT can be as long as several months or years. In order to effectively reduce the test time, we often use the Finite Element Method (FEM) instead of TCT.Although FEM takes less time than TCT, it still takes some time to use the FEM to get the simulation result. We build the model according to different parameters and fixed boundary conditions, and then apply the thermal cycle load on the model to obtain the prediction life value of the electronic package. Different researchers may lead to different results, even if they use the same model parameter. If we use a large amount of validated FEM data to build a database for machine learning(ML), then we can immediately evaluate the electronic package prediction life through machine learning methods. It not only saves the time to build the model and validation, but also avoid the simulation error.This research is to use the ML method to analyze the reliability of Wafer Level Chip Scale Packaging (WLCSP). ML can find potential rules of data sets through algorithms, establish mathematical models, and obtain prediction life of different WLCSP structures. For the same data set, different algorithms may get different results. So, choosing the suitable algorithm is the most important step in using machine learning algorithm. This research uses a Random Forest algorithm (RF) and Extremely Randomized Trees (ET) algorithm to evaluate the reliability of WLCSP. The training database was generated by FEM and compared with the TCT experimental results to verify the FEM model. After obtaining the verified FEM model, in the same modeling process, we design feature levels to generate multiple data sets with different data volumes and different data distributions. Discuss the impact of data volume and data distribution on the RF model and ET model.