Utilizing CNN to predict homogeneous thermo-mechanical properties of conductive layers for reliability numerical analysis in electronics

Abstract

This study explores the application of Convolutional Neural Networks (CNN) in predicting the partitioned homogeneous properties (PHPs) of electronic product wiring structures, aiming to enhance the efficiency of reliability analysis through Finite Element Analysis (FEA). A systematic and novel method was developed to generate the input partitioned wiring diagram image sets that foster model generalization and universality. The performance of the CNN-based approach was assessed through regression analysis by employing a leave-one-out cross-validation (LOOCV) approach across three PCBs. Additionally, the predicted PHPs of one of the PCBs were applied in product-level FEA to validate their reliability, further demonstrating the practical applicability of the CNN-based method. The results demonstrate that a well-trained CNN model can accurately predict the properties of previously unencountered wiring structures, thereby facilitating direct application in product-level reliability FEA and improving the efficiency of reliability assessment. Furthermore, efficiency evaluation revealed that the CNN-based method offers significant advantages in terms of time and cost economy compared to the mesoscopic FEA method in determining, highlighting its potential for broader application in electronic product reliability analysis. The findings provide preliminary insights and propose strategies to enhance the applicability of CNN methods in this domain, ultimately aiming to improve the efficiency and reduce costs in reliability assessments, thereby streamlining the overall product development process.