Feasibility Investigation of Machine Learning for Electronic Reliability Analysis using FEA

In this paper, a machine learning framework and model is presented for predicting the vibration reliability of solder joints using finite element analysis. Though some models covering temperature cycling conditions are known, mechanical loads like vibration and shock are rarely covered. The proposed model predicts the elastic strains of the solder joints under isothermal harmonic vibration conditions based on temperature, amplitude and design data, e.g. chip thickness and solder joint diameter. The finite element model used for data synthesis is based on an already established specimen design specifically dedicated to investigate Flip-Chip components under combined vibration and TC loads. Harmonic analyses were carried out using a quarter model of the specimen. Equivalent elastic strain was extracted following a dynamic selection of the highest stressed elements per solder joint. A standard feed-forward artificial neural network was used due to its simplicity to keep overall complexity and computational efforts low. The machine learning framework was investigated in terms of model complexity, performance and required training data. Additionally, its resilience against noise, typically present by acquiring experimental data, was studied. Since all data sets were obtained using finite element analyses, artificial Gaussian noise is added to the data sets to strengthen the models robustness. The prediction results show that the model achieves high accuracy (MSE ≥ 2.73*10
⁻⁵
) while also keeping the computational time low. Since only one output feature is used, a less complex framework can be used to accurately forecast the equivalent elastic strain.