Validating Junction Temperature Estimation of Power Modules using Generative AI and Fiber Optic Sensing

Accurate junction temperature <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(T_{\mathrm{j}})$</tex> estimation in switching power devices is crucial for the reliability and efficiency of power electronic systems. Traditional thermal models, based on analytical simulations, often fail to capture non-uniform heat distribution and transient behaviour during load changes, resulting in inaccurate estimations. To address this challenge, this paper proposes a generative AI approach which is trained from a linear time-invariant RC thermal network with minimal data inputs to form a temperature-dependent thermal (TDT) RC model. This model aims at capturing the temperature dependence of each layer in the stack-up of the power semiconductor dies, and the impact of thermal cross-talk in multi-die arrangements, as seen in high-power modules. Simulations are conducted to demonstrate the feasibility of the proposed approach, and the estimated <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$T_{\mathrm{j}}$</tex> is compared with conventional RC thermal models. Additionally, to establish the accuracy of the proposed model in real-world scenarios, a validation methodology is presented, comparing the proposed model with fiber-optic temperature sensing for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$T_{\mathrm{j}}$</tex> estimation. A comparison of estimated temperature between the AI-based RC thermal model and experimental values show a relative error of 2.17% for the central die and 4.52% for the peripheral die, highlighting the capability of the proposed methodology to accurately capture non-uniform heat distribution and transient thermal behaviour. Moreover, the high estimation bandwidth of 20.03 Hz enables the model to track rapid temperature changes during transient events, ensuring accurate <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$T_{\mathrm{j}}$</tex> prediction under dynamically varying power losses.