Predicting Effective Thermal Conductivity of Sintered Silver by Microstructural-Simulation-Based Machine Learning

Effective thermal conductivity (ETC) is an important property of sintered Ag for its application as die-attach material. However, accurate and fast ETC evaluation is still challenging for the conventional approaches. In this study, microstructural-simulation-based machine learning was performed to predict the ETC of sintered Ag. The microstructure-based finite element (FE) model was established by converting each pixel of microstructural image into one element, which saves the computational time while maintaining excellent prediction accuracy (relative error of 4.3% with the measurement result). The FE-simulated ETC values were then applied as target values of the machine-learning database and four manually selected microstructural descriptors (porosity, shape factor, dominant heat-transfer path and major heat-transfer barrier), which quantitatively describe the pore features of sintered Ag, served as the inputs. The trained machine-learning model, i.e., support vector regression model (whose hyper-parameters were optimized by genetic algorithm), can achieve efficient and accurate ETC prediction (determination coefficient R2 of 0.974), which significantly outperforms the existing analytical (severe overestimation) and semi-empirical models (R2 of 0.880 and 0.947). This work demonstrates the superiority of machine learning method and pioneers a pathway for ETC prediction, which can also be further extended to investigate other effective properties of sintered Ag.