Highly accurate and efficient prediction of effective thermal conductivity of sintered silver based on deep learning method

• Microstructure-based FE simulation was implemented to evaluate ETC of sintered Ag. • Database was created with actual microstructures and reliable FE simulated values. • The trained CNN can predict ETC of sintered Ag both accurately and efficiently. • Prediction accuracy of CNN outperforms analytical and machine learning models. • The model trained by sintered Ag dataset can be used to predict ETC of sintered Cu. Effective thermal conductivity (ETC) of sintered Ag is an essential parameter for its die-attach application in power electronics packaging, which could vary significantly with sintering conditions. The existing ETC evaluation approaches are either of limited accuracy (analytical methods) or resource- and time-consuming (experiments and numerical simulations). In this study, deep learning method based on convolutional neural network (CNN) was first performed to predict the ETC of sintered Ag. The database was created with 6156 realistic microstructures of sintered Ag and corresponding reliable microstructural finite-element simulated ETC values (relative error of 5% with experimental results). Based on the appropriate design of CNN architecture, the trained model can accurately predict ETC values of the testing dataset samples with determination-coefficient (R 2 ) of 0.987, which significantly outperforms the conventional analytical (R 2 of 0.837) and machine learning methods (R 2 of 0.951). Besides, the prediction by CNN takes merely 0.14 s for an image, which is almost negligible. The methods presented here open a new way to achieve highly accurate and efficient prediction of ETC, which can help to prepare sintered-Ag die-attachment with desired ETC for power devices and also be applicable to investigate other effective-properties of sintered Ag.