With the emerging technologies like autonomous driving and x-by-wire systems, the vehicle's onboard power supply system, also known as the powernet, is subject to stringent safety requirements. Failure of the powernet leads immediately to the loss of all the safety-related functions such as braking, steering, autonomous driving features, etc. Among all the powernet components, special attention shall be paid on batteries due to their complex electrochemical nature. Accurate battery diagnosis is essential to predict the performance of batteries. Safety validation is, therefore, required to guarantee that the prediction deviations is acceptably low under all real-world conditions. To ensure the robustness of battery diagnosis under real-world conditions, uncertainty quantification (UQ) is necessary to assess the prediction deviations. This thesis aims to develop a probabilistic surrogate model using generative AI models (e.g., Conditional Generative Adversarial Networks (cGAN), Conditional Normalizing Flows (cNF) to quantify and distinguish different sources of uncertainty. Additionally, the study will explore strategies to estimate failure probabilities with a limited amount of data while maintaining high accuracy. The research questions are: how to quantify and separate epistemic and aleatory uncertainties, which generative AI models are best suited for capturing aleatory uncertainties, how to evaluate and calibrate the reliability of the predicted uncertainty, how does the amount of available data impact the accuracy of uncertainty quantification, how to estimate failure probability effectively with a small data set.

• As part of your Master thesis, you will conduct an extensive review of the state-of-the-art methodologies for uncertainty quantification (UQ) based on generative AI.
• You will prepare synthetic data for study.
• Furthermore, you will implement and compare different UQ models in terms of their ability to accurately capture predictive uncertainty and to distinguish between epistemic and aleatory uncertainty.
• In addition, you will improve the prediction interval calibration to ensure a robust and reliable uncertainty estimate.
• Last but not least, you will develop an approach to estimate the probability of battery diagnostic failure with minimal data using probabilistic surrogate models.

• Education: Master studies in any field
• Experience and Knowledge: background in Machine Learning, Deep Learning, Reliability or Uncertainty quantification, etc.; experience with PyTorch and deep generative models (cGANs, cNF, etc.); knowledge of battery diagnostics or battery systems is a plus
• Personality and Working Practice: you are a self-motivated and proactive person who is able to work independently
• Languages: very good communication skills in written and spoken German or English