Data-driven optimization for enhanced thermal management of 3D stacked chips

Three-dimensional (3D) stacked chips, with their exceptionally high power densities, face significant thermal challenges. However, traditional design methods struggle to address the complex coupled variables involved in system-level optimization. In this study, we propose a data-driven optimization framework that combines thermal simulations with machine learning to efficiently navigate and refine the chip thermal design space of 24 variables. By employing Latin hypercube sampling, we generate design solutions that are evaluated using an automated simulation pipeline. These solutions are then used to train surrogate models built from different machine learning algorithms. The surrogate models drive high-throughput screening, combined with thermal simulation verification, resulting in substantial improvements in heat dissipation compared with the initial dataset: the equivalent thermal resistance distribution is narrowed from 1.18 – 60.75 to 0.63–2.81 K W −1 , reducing the range by 96% and the minimum by 47%. Additionally, interpretable analysis quantifies the contributions of each variable to heat dissipation. Our results present a scalable and effective framework for thermal management in 3D integrated systems, paving the way for the development of more reliable, high-performance chips.