High-Efficiency Extraction Method for Thermal Network Model of Advanced Electronic Packages

Abstract

To reduce the number of grids in numerical computations that involve multiple length scales, a compact thermal model of an electronic package is widely used for electronic system thermal design. This study was conducted according to development of libraries of physical models for an integrated design environment (DELPHI) procedure and extended to dual-chip application. A flip-chip chip-scale package (FCCSP) containing a single or dual chip was chosen as the test vehicle. A domain-knowledge-assisted sequential least-squares programming (SLSQP) technique is used to minimize the objective function; meanwhile, the genetic algorithm (GA) technique is used for comparison. This study examines how to use domain knowledge on initial condition settings when using SLSQP and how to increase spatial resolution to improve the network models effectively. It is found that using calculated 1-D thermal resistances as the initial estimates of the critical thermal resistance parameters can increase the accuracy of the network model training by SLSQP, which has accuracy similar to GA and much better computational efficiency. This study also demonstrates that the network model can be improved by systematically increasing its spatial resolution of the main heat dissipation path based on the errors observed in primary attempts, which has more benefits than increasing the spatial resolution of the surface having a larger temperature difference.