Physics-based Nested-ANN Approach for Fan-Out Wafer-Level Package Reliability Prediction

For the current electronic package simulation evaluation process, the main steps include material characterization, data treatment and finite element modeling (FEM). In the workflow, the data treatment and FEM computation are complex and time-consuming, especially when time-and temperature-dependent behaviors are involved. In this work, a physics-based machine learning (PB-ML) nested-artificial neural network (ANN) algorithm is developed to optimize the process flow and improve work efficiency. The proposed physics-based nested-ANN model consists of a material-ANN and a mechanical-ANN. The material-ANN is the initial surrogate model constructed and trained by raw test data and physical domain to generate material property data for further numerical computation, while the mechanical-ANN is a sequent surrogate model implementing the output of the material-ANN as a feature and the FEM data as a label to train and predict the package system performance. The proposed framework is developed and applied to a fan-out wafer-level package (FOWLP) reliability evaluation considering the effect of viscoelastic property of epoxy molding compound (EMC) materials. The system performance of the FOWLP during the post-molding cure (PMC) process is characterized by the framework. The material-ANN employs dynamic mechanical analysis (DMA) data of a series of EMC materials as input and mathematic models, including the Prony model and WLF formula, to characterize the viscoelastic behavior as outputs, which are then forwarded to the mechanical-ANN as inputs to predict the mechanical response of the package during PMC. The average maximum accurate percentage error (MAPE) and the coefficient of determination (R ² ) are 8.87%, 0.7793, 4.99%, 0.9545, 14.89% and 0.8861 for the maximum accumulated creep strain (CEEQ), the maximum von Mises stress and the maximum warpage deformation, respectively. The well-constructed nested-ANN model demonstrates a calculation time of 13.2s which is highly reduced compared with the data treatment and FEM process. This method is proved to be featured with high efficiency inherited from PB-ML model and accuracy from physical governing functions. The framework developed in this work paves a new way for computation source saving and work process acceleration in packaging system reliability evaluation.