GA-NN and Explainable AI for IC Packaging Warpage Optimization: A Case Study on Product Feature

In semiconductor packaging, avoiding product warpage is difficult since it involves product specifications and related processes. In order to effectively predict warpage, traditional methods use a large amount of data to perform statistical analysis. However, in the actual field, obtaining detailed parameter data for each process requires a lot of time and manpower, which makes statistical method impractical. In other words, achieving accurate warpage prediction with limited data is crucial. Therefore, this research proposes an artificial neural network model with genetic algorithms to enhance warpage prediction. The proposed algorithm uses genetic algorithms to achieve optimize the hyperparameter of artificial neural network (ANN), which is called Genetic Algorithm Optimized Artificial Neural Network (GA-NN), the algorithm can accelerate the convergence speed and performance of the ANN. The data set was generated by the finite element method from the semiconductor packaging product feature data. Compared with other algorithms, the product data simulated by the proposed method can maintain numerical stability and achieve rapid convergence. In the proposed ANN, an interpretable AI algorithm - SHapley Additive exPlanations (SHAP) is also used to further analyze various input parameters and quantify the operation results of the machine learning model. The empirical result shows that the prediction has an error of only 3 % on Molding Functional Chip Ball Grid Array (MFCBGA) package products. Compared with traditional algorithms and other machine learning methods, the proposed algorithm performs faster and with higher prediction accuracy, while providing product feature influence and visualizes the direct relationship between product specifications and warpage. To sum up, this research achieves effective package warpage prediction, while further explains the relationship between warpage generation in the AI model, input data and prediction targets to customers. Finally, the product design cycle can be reduced and the product quality can be improved simultaneously.