A hybrid modeling framework for predicting the multilayer ceramic capacitor reliability under thermal-electrical coupling operating conditions

Physics-of-failure (POF) methods based on testing are largely inadequate in meeting the reliability prediction demands of large-scale production. In contrast, pure data-driven methods depend heavily on sample quality and completeness, resulting in weak extrapolation and unclear mechanisms. This paper presents a modeling framework that synergizes POF with data-driven methods, combining their scalability in extrapolation and accuracy to address the shortcomings of reliability prediction for multilayer ceramic capacitors (MLCCs) under thermo-electric coupling (TEC) conditions. A performance-distributed degradation POF model is refined for MLCCs, considering both operating conditions and manufacturing parameters, and a breakdown reliability POF model is developed that incorporates the Schottky barrier and Weibull distribution. Furthermore, the neural networks are enabled to compensate for the prediction errors of the POF model by capturing the operating conditions, manufacturing parameters, and time-series features. Additionally, Bayesian optimization (BO) is employed to optimize the hyperparameters, thereby enhancing the prediction accuracy with limited sample sizes. The root mean square error (RMSE) of the hybrid model decreases by up to 82.78% compared to the POF model. Finally, the modeling framework demonstrated its potential through a 2000-h extrapolation proof-of-concept, which used a new MLCC type with the same structure and studied material.