The effect of data distribution in Ensemble Learning Algorithms on WLCSP reliability Prediction

As the consumer market is growing faster and faster, the technology is changing with each passing day, and we are focusing much more on the reliability of electronic packaging. Electronic packaging is primarily used to protect the complex structure of IC components, which is of considerable importance in the semiconductor industry. We are working on assessing the reliability of electronic packaging using Accelerated Thermal Cycle Testing (ATCT). ATCT is an important test to evaluate the reliability of packaging. ATCT can judge the reliability of electronic packaging, but the experiment requires a many time and expense. It commonly applies the Finite Element Analysis (FEA) replace of experiments to efficiently decrease the testing time. Although FEA spends less time than ATCT, using FEA takes a lot of time building a model to obtain simulation results. We build models based on different parameter settings and apply thermal cycling loads on the models to get the reliability life of electronic packaging. However, multiple researchers may also get different simulation results. With the development of computer equipment, computing performance has high performance and with the booming development of artificial intelligence, this research applies machine learning methods. If we can use a validated Finite Element Model (FEM) data to generate a series of the data for Machine Learning (ML). It can immediately apply ML methods to judge the reliability of electronic packaging. It can save time to build models and avoids the mistake in simulation. Four algorithms are used in this study to evaluate the reliability of Wafer Level Chip Scale Packaging (WLCSP). These algorithms are Random Forest, Extremely Randomized Trees, Adaptive Boosting, Gradient Boosting. We verify our simulation results by comparing them to experimental results. When the simulation of the model is successfully verified, we will use the same modeling method to build a different database with multiple data volumes from FEM. To explore the influence of the distribution of different data volumes on these algorithms and find the algorithm with the most stable prediction performance and performance results.