Modeling of Heterogeneously Integrated Systems: Challenges and Strategies for Rapid Design Exploration

Design space exploration in Heterogeneously Integrated (HI) systems requires trade-off between electrical, thermal, and mechanical performance metrics. Therefore, Multiphysics models of high fidelity are necessary. But, the models also need to span multiple length scales connecting behaviors at sub-micrometer features to response at centimeter sized modules. Recently machine learning models are proposed as enabling rapid design exploration, but their fidelity to the partial differential equation solution surface that underlie the physical behavior remains to be proven. In this talk we systematically explore (1) adaptive grid solution strategies for systematic trade-off between solution accuracy and computational speed, (2) decomposition of problem domain to enable compact models of sub-domains and to enable a coordinated multi-level analysis, (3) machine learning models as compact models of decomposed domains, and (4) the computational cost of training and using machine learning models relative to physical models.