Electro-Thermal Modeling and Surge Reliability of High-Power Semiconductor Devices with Machine Learning Based Parameter Prediction

High-power semiconductor devices are considered key components of modern power electronic systems operating in high-voltage and high-current applications such as high-voltage direct current transmission, railway traction, industrial power conversion, and electric vehicle fast-charging infrastructure. As power density continues to increase, the reliable operation of these devices is increasingly constrained by thermal stress and transient overload conditions, particularly during surge current events. Therefore, it is a critical challenge to accurately estimate junction temperature and assess the reliability of device robustness. The rapid increase in junction temperature influences the reliability of the device, both at the device design and system levels, leading to performance degradation and device failure. This thesis proposes electro-thermal and data-driven modeling methodologies to address these challenges, with a primary focus on diodes and thyristors. First, the transient thermal impedance of high-power press-pack diodes is characterized by the development of an electro-thermal modeling framework. Thermal resistance and capacitance parameters are analytically derived from device geometry and material properties, enabling the construction of Cauer thermal networks that are subsequently transformed into equivalent Foster thermal models. Moreover, curve-fitting techniques are performed in MATLAB® to extract thermal parameters from transient thermal impedance responses. The proposed models are validated through experimental measurements. Building upon this foundation, a model is proposed to estimate the forward voltage drop for high-power diodes under high surge current levels. Moreover, simulations were also performed to estimate the junction temperature. By integrating rated electrical parameters with thermal simulation data, the proposed methodology captures dynamic slope resistance and transient voltage behavior to counter the limitations of conventional static current-voltage based temperature estimation methods. Moreover, the results of the proposed model are compared with experimental results. In addition, a comparative study of two different packaging assemblies is conducted to investigate and analyze the thermal performance and surge capability. An insulated module diode and a press-pack diode with pressure contact technology (PCT), both with identical specifications and power ratings, are utilized to evaluate the influence of packaging assembly on device performance. For this purpose, a dedicated experimental setup was developed, generating a 10 ms half sine wave, and a series of tests of the forward voltage drop across varying surge levels were conducted. Moreover, electro-thermal simulations are also performed to analyze and compare junction temperature and analyze failure behavior. In addition to physics-based modeling, this thesis introduces machine learning (ML) frameworks for classifying and predicting key design and performance parameters, including chip diameter and average current of high-power semiconductor devices. A comprehensive survey of high-power thyristors is conducted, analyzing the influence of chip diameter and thickness on both electrical and thermal performance. Moreover, a dedicated dataset is developed by extracting electrical parameters from the leading semiconductor manufacturer’s datasheet of multiple models. Moreover, multiple ML models, including gaussian process regression (GPR), support vector machine (SVM), artificial neural network (ANN), and ensemble methods, are also implemented. Furthermore, data augmentation, shapley additive explanation (SHAP) and feature optimization are also performed. Overall, this research establishes a unified framework, combining electro-thermal modeling, simulation, experimental validation, and data-driven techniques to improve the reliability assessment and design of high-power semiconductor devices. The proposed methodologies provide practical tools for manufacturers and system designers, contributing to the development of robust and reliable devices and systems.