Logo image
Mechanical reliability of high-power modules via simulation-based machine learning
期刊文章

Mechanical reliability of high-power modules via simulation-based machine learning

C.-C. Lee, W.-C. Tsai 和 J.-C. Chuang
Engineering Applications of Artificial Intelligence, 卷.154
2025
Web of Science ID: WOS:001487360200001

摘要

Finite element analysis (FEA) High-power module packaging Machine learning (ML) Process-induced warpage Support vector regression (SVR) Thermal cycle reliability Finite element analyse Finite element analyze High power module High-power module packaging Machine learning Machine-learning Module packaging Process-induced warpage Support vector regression Support vector regressions Thermal cycle reliability Warpages Support vector machines
The thermal cycle reliability and process-induced warpage related to high-power module packaging integrated with the type of insulated gate bipolar transistor modules that have direct bonded copper are critical factors remarkably influencing the longevity and performance of power electronics. Despite the robustness of traditional finite element analysis (FEA) methods, they have profound limitations, including high computational demands and the necessity for extensive experimental validation. This study presents an integrated approach that combines FEA with machine learning (ML) to improve predictive reliability. The ML model integrated support vector algorithms is employed to trained base on FEA-generated stress, strain, and warpage data to predict outcomes with better accuracy and mitigated computational overhead. Hyperparameter optimization through grid search improves the model's predictive capabilities. The proposed ML-FEA framework, validated through comparison with experimental results, effectively forecasts failure probabilities and process-induced warpage under diverse thermal cycling scenarios. This integrated approach improves predictive accuracy and offers a scalable solution for the reliability assessment and design optimization of high-power electronic modules. Accordingly, it provides a foundational reference for mitigating warpage and enhancing overall reliability. Due to predictive accuracy, R2 value is employed to check the ML prediction. In this study, R2 value for ML in training data is more than 0.95. In addition, for computational time decreasing, the time from 604,800 s to lower than 60 s is mitigated. © 2025 Elsevier Ltd

檔案與連結 (1)

url
https://www.scopus.com/inward/record.uri?eid=2-s2.0-105004179581&doi=10.1016%2fj.engappai.2025.111019&partnerID=40&md5=5827d8fb29aeccbbad7d4b10804e86f2檢視

相關連結

指標

1 檢視次數

詳細資料

Logo image