Warpage prediction of fan-out wafer-level package based on coupled deep learning and finite element simulation

In recent years, fan-out wafer-level package (FOWLP) has gained widespread attention in integrated circuit industry due to its significant potential in enhancing packaging performance, reducing costs and minimizing size. However, accurate prediction of warpage in FOWLP remains a formidable challenge, as conventional prediction methods often suffer from prolonged iterative cycles and high computational costs. This study integrates the finite element method (FEM) with artificial intelligence (AI) techniques to develop highly accurate and efficient warpage prediction model for FOWLP based on varying chip sizes and spacings. Automated modeling and data generation were performed using simulation technique and python scripts, resulting in the creation of two datasets of different scales for regression training and optimization. Prediction results indicate that Residual Network-152 (ResNet-152) performs best on smaller datasets, while Global Context Vision Transformer-Tiny (GCViT-Tiny) exhibits greater stability on larger datasets. The Huber loss function was employed to optimize the deep learning (DL) model weights through backpropagation , significantly improving both training efficiency and prediction accuracy. Additionally, the reliability and practicality of the DL model were validated by performing simulations and AI predictions on structures not included in the dataset. Finally, the trained DL model provided brief technical guidance for optimizing warpage in wafers for the integrated circuit industry. In terms of efficiency, DL models offer a clear advantage in industrial applications. The research results provide effective theoretical support and practical guidance for FOWLP optimum design and reliability assessment, demonstrating significant application potential. • Proposed an automated modeling and simulation script for FOWLP • Established a deep learning model for FOWLP warpage prediction • Achieved deep integration between the FEM and AI • Evaluated and optimized multiple visual neural network architectures • Validated AI model reliability and provided technical guidance for industry