MACHINE LEARNING OPTIMIZATION OF UNDERFILL FLOW TIME IN FLIP-CHIP ENCAPSULATION OF BGA ASSEMBLIES

Predicting and minimizing underfill flow time remains challenging due to complex nonlinear interactions among material properties, solder bump geometry, and dispensing strategy. This study presents a hybrid computational framework integrating computational fluid dynamics (CFD) simulations with machine learning–based predictive modeling to analyze and optimize underfill flow behavior. A total of 120 simulations were performed using ANSYS Fluent, covering six dispensing configurations, four solder bump geometries (spherical, cylindrical, hourglass, and concave), and five underfill materials with varying viscosity, surface tension, and density. To eliminate volumetric bias, an inlet-area-balanced dispensing strategy was implemented to maintain constant total injection volume across configurations. Taguchi-based parametric analysis revealed that underfill material viscosity is the dominant factor governing flow time, followed by dispensing configuration, while bump geometry shows comparatively minor influence. Bayesian optimization using Scikit-Optimize and Optuna identified the optimal configuration as hourglass-shaped solder bumps, U-33.3 three-inlet dispensing, and a low-viscosity methanol-based underfill. Among the evaluated machine learning models, Random Forest achieved the highest predictive accuracy (RMSE = 1.19), demonstrating strong capability in modeling nonlinear process interactions and enabling data-driven optimization of encapsulation processes. The CFD–ML framework provides a computationally efficient tool for encapsulation process optimization, enabling accelerated virtual prototyping and data-driven design in microelectronic packaging.