End-to-end Design Automation for Reliable Power System Components

To address the aforementioned challenges we introduce an end-to-end methodology for providing good-quality data generated or sampled from targeted regions of the parameter space.

This method combines generative modeling, representation learning, surrogate modeling and optimization techniques to create a flexible framework applicable across diverse power system component design domains and establish its merits with an application to automated design of power semiconductor devices. The approach incorporates reliability-driven constraints and multi-physics considerations to ensure designs remain robust against real-world operational uncertainties and enhances standard optimization pipelines with targeted local improvements.

While traditional optimization methods often operate directly in the full-parameter space, our approach introduces a local design enhancement strategy which is instrumental in power systems engineering applications as it contributes to efficient computational resource allocation. We evaluate the merit of the proposed end-to-end framework for design automation with a case study of Insulated Gate Bipolar Transistor (IGBT) termination design. Moreover, we describe the problem setting for two other use cases in the proposed framework to showcase the possible extension of embodiments of the framework in engineering design automation tasks for power system assets.