Machine Learning Enabled Optimization of Pick-up Process for Thin Die

Driven by the demand of artificial intelligent, high-performance computing, electric vehicle and smart city, semiconductor industry is building high performance and high integrated systems. To increase the energy efficiency of these complicated systems, particularly in power modules, reducing the power consumption is needed. Thinning the die is one of the solutions, which can provide a lot of advantages including faster heat dissipation, low junction temperature, low electrical resistance, etc. However, due to its sensitivity of stress, processing of thin die is a challenging, including die pick-up process. In this paper, we proposed a methodology using machine learning and finite element method (FEM) to optimize the pin designs with the objective to minimize the stress in the die during pick-up process. A dynamic numerical model was first built up to simulate the stress distribution in the die during pick-up process, and a Gaussian process regression was applied to generate the relationships between pin parameters and die stress. Meanwhile, a Bayesian optimization was used to optimize the parameters. The optimized pin designs using the above methodology was further validated by FEM. Compared with traditional DOE method, the proposed methodology using machine learning and FEM exhibits much higher effectiveness in parameter optimization with less testing cost and efforts.