Solder Joint Reliability Risk Estimation by AI-Assisted Simulation Framework with Genetic Algorithm to Optimize the Initial Parameters for AI Models

Abstract

Solder joint fatigue is one of the critical failure modes in ball-grid array packaging. Because the reliability test is time-consuming and geometrical/material nonlinearities are required for the physics-driven model, the AI-assisted simulation framework is developed to establish the risk estimation capability against the design and process parameters. Due to the time-dependent and nonlinear characteristics of the solder joint fatigue failure, this research follows the AI-assisted simulation framework and builds the non-sequential artificial neural network (ANN) and sequential recurrent neural network (RNN) architectures. Both are investigated to understand their capability of abstracting the time-dependent solder joint fatigue knowledge from the dataset. Moreover, this research applies the genetic algorithm (GA) optimization to decrease the influence of the initial guessings, including the weightings and bias of the neural network architectures. In this research, two GA optimizers are developed, including the “back-to-original” and “progressing” ones. Moreover, we apply the principal component analysis (PCA) to the GA optimization results to obtain the PCA gene. The prediction error of all neural network models is within 0.15% under GA optimized PCA gene. There is no clear statistical evidence that RNN is better than ANN in the wafer level chip-scaled packaging (WLCSP) solder joint reliability risk estimation when the GA optimizer is applied to minimize the impact of the initial AI model. Hence, a stable optimization with a broad design domain can be realized by an ANN model with a faster training speed than RNN, even though solder fatigue is a time-dependent mechanical behavior.

Development of ANN-Based Warpage Prediction Model for FCCSP via Subdomain Sampling and Taguchi Hyperparameter Optimization

“…In addition, the optimal values of the hyperparameters are highly problem-dependent. In this investigation, the initial/untuned hyperparameter values, namely, optimizer (Adam) [ 27 ], activation function (ReLU) [ 15 ], the number of hidden layers (three) [ 28 ], and the estimate of the upper limit of the number of neurons in each layer [ 29 ], are specified based on the literature results. Note that the size of the training dataset is directly related to the number of neurons applied in each layer.…”

“…To solve the aforementioned challenges and even lessen the prediction uncertainty and modeling error made by less experienced engineers, researchers start seeking the integration of simulation and machine learning (see, e.g., [ 11 , 12 , 13 , 14 , 15 , 16 ]). To date, due to the rapid advance of computer technologies and machine learning algorithms, it has evolved into a critical tool for addressing a wide range of real-world issues, with applications covering medical diagnosis, transportation, space exploration, defense systems and various engineering fields.…”

“…This has allowed the development of a deep learning-based prediction model for timely and effective predictive analysis of very complex systems. The simulation-based deep learning prediction models have been extensively applied in advanced microelectronic packaging for a quick and accurate assessment of their thermal-mechanical performance, such as thermal performance [ 12 , 13 ] and reliability [ 14 , 15 , 16 ]. For example, Law et al [ 12 ] developed a deep learning model for the prediction of the thermal performance of quad flat no-lead (QFN) packages using an ANN.…”

“…Subbarayan et al [ 14 ] applied an artificial NN (ANN) algorithm to build up a reliability prediction model for a ball grid array (BGA) package. Yuan et al [ 15 ] applied an ANN-based simulation framework to investigate the solder joint reliability of a wafer-level chip-scale package, where the initial parameters of the ANN model, namely, the weights and bias, were obtained using a genetic algorithm (GA). Hsiao and Chiang [ 16 ] combined FEA together with random forest (RF) to explore the solder joint reliability of wafer-level packaging subjected to thermal cycling.…”

Deep learning-based methods in structural reliability analysis: a review

“…The validity of their model is verified with sensor data recorded from gas turbine structures. An approach similar to [55] is also taken by Yuan et al [68]; they used RNNs for solder joint reliability after fatigue loading. Their research follows the AI-assisted simulation framework and builds the non-sequential ANN and sequential RNN architectures to deal with the time-dependent and nonlinear characteristics of the solder joint fatigue failure.…”