Data-driven optimization scheme for uniform jet cooling

Recent advancements in semiconductor packaging technologies, such as chiplet technology, 3D stacking, and heterogeneous integration, have significantly enhanced device performance. However, these next-generation packaging techniques can introduce severe thermal issues that adversely affect device reliability. The integration of diverse components with varying materials and operating powers within a single package can result in substantial surface temperature gradients on the substrate. This leads to thermal coupling effect, where heat propagates from local hotspots with high heat flux to temperature-sensitive components (i.e. memory), potentially degrading overall package performance. Additionally, differences in thermal expansion coefficients within materials may cause physical and permanent damage, such as thermal crack. To address these challenges, appropriate thermal management strategies that ensure uniform temperature distribution across the package are essential. Multiple jet impingement cooling emerges as a promising approach, offering high heat transfer coefficients and the ability to enhance surface temperature uniformity through adjustments in jet nozzle arrangement. We identify the limitations of traditional jet nozzle arrays with uniform jet spacing in achieving temperature uniformity and aim to optimize the nozzle arrangement that can improve temperature uniformity without additional energy consumption. To efficiently achieve this, a numerical data-driven surrogate model based on convolutional neural networks (CNN) was developed, supported by an active learning strategy. Moreover, our proposed hierarchical exploration algorithm has shown excellent optimization efficiency in identifying superior jet nozzle arrays within a vast design space.