An active learning approach to performance optimization of jet array impingement-based cooling module for heterogeneous integration semiconductor packages

Abstract

The heterogeneous integration package, gaining attention as a next-generation packaging technology, offers high design freedom but can lead to serious issues like thermal crack, delamination, and performance degradation due to severe temperature gradients caused by non-uniform heat flux distribution. In this thesis, a direct liquid cooling module using multiple jet impingement arrays was designed, and the optimal nozzle arrangement was explored to minimize temperature non-uniformity of package substrate with non-uniform heat flux distribution and energy consumption for cooling. For this purpose, the multi-objective optimization based on genetic algorithms was conducted using finite element method-based numerical analysis model and convolutional neural network metamodel. Additionally, a design optimization framework, combining hierarchical exploration and active learning, was developed to reduce the time and resources required for optimization. It was confirmed that the optimized nozzle arrangements for three cases exhibited superior performance compared to intuitively arranged regular nozzle arrays.