Artificial Intelligence-Based Warpage Prediction Model for Accelerating Thermo-Mechanical Simulation in Advanced Packaging

One of the most challenging issues in advanced packaging is the increasing design complexity, which necessitates the expenditure of more effort in finding an optimal design, compared with traditional packaging methods. The use of multiple chiplets with 3D stacked dies and a number of passive components increases the dimensionality of design variables, making finite element thermomechanical simulations more computationally demanding. Moreover, as the variety of materials used in advanced packaging increases, coefficient of thermal expansion (CTE) mismatches are more likely to occur and lead to geometrically uneven warpage. Given these challenges, a machine learning-based warpage prediction model can be adopted to accelerate the design process. A well-trained model can provide immediate results without solving iterative and nonlinear governing equations while also handling high-dimensional design optimization problems. This study presents a machine learning-based warpage prediction model that shows less computation time, compared with thermomechanical simulation. Considering the uneven warpage characteristics of advanced packaging, we developed a conditional generative adversarial network (cGAN)-based model capable of predicting the global warpage distribution across the entire package surface. The training dataset for the cGAN model was generated from a warpage simulation model of a commercial 2.5D package with multiple chiplets, high bandwidth memory, and a Si-bridge interposer. The thermomechanical properties of the epoxy molding compound, including the CTE, thickness of the Sibridge interposer, and process temperature, were considered as key design variables influencing warpage. The warpage distribution predicted by the cGAN model was evaluated against actual thermomechanical simulation results under the same conditions. The developed cGAN model enables rapid warpage prediction and design space exploration, which reduces the total computation time, compared with repetitive thermomechanical simulation. This study concludes by discussing future applications of the presented model for the acceleration of global warpage optimization in advanced packaging, including experimental validation.