Rapid Estimation of Anisotropic Thermal Conductivity in RDL for 2.5D Chiplet Design

This paper presents a framework for rapidly estimating the anisotropic thermal conductivity for redistribution layers (RDL) in 2.5D chiplet integration. Most current approaches for thermal property evaluation requires detailed routing information, which is time-consuming to generate and typically unavailable during early design stages. We propose a novel machine learning strategy that predicts anisotropic thermal properties using only copper density distribution data, which can be quickly obtained from the global routing information. Furthermore, this approach is resolution-independent, allowing the model to be trained on patches of one size while predict thermal properties for patches of different sizes. This enables thermal evaluation to occur concurrently with 2.5D advanced package design, facilitating early-stage optimization for heterogeneous packaging and supporting a broader range of design scales.