Aplicații de învățare automată în proiectarea tranzistoarelor MOSFET de putere / Machine Learning applications in power MOSFET design

Design optimization plays a crucial role in advanced discrete device design, as the performance requirements are constantly increasing. For MOSFETS, there is a need for high accuracy in on-state resistance Ron simulation. To achieve high accuracy, there is a requirement for a complex 3D finite element simulation, which usually takes tens of minutes. Furthermore, preparing the simulation (e.g., model setup, adjusting design parameters, solving numerical issues, etc.) additionally increases the time for obtaining results. Therefore, designing new devices involves a lot of iterative, time-consuming manual work. In this work, we introduce an automatic design optimization methodology for MOSFETs that minimizes the amount of manual work of engineers and the design time of new devices. First, we propose methods for reducing the necessary time for analysis of the device by using an automated simulation setup and further, alternative methods for simulation. Hence, a high accuracy prediction metamodel (error less than 2%) is obtained using only a couple hundred 3D finite element simulations for training. Secondly, we propose methods for manufacturing high-performance devices at the lowest cost, which require an optimal design with regard to both material loss and yield. The optimal design parameters of any device technology, which need to meet an expected value of Ron, will be automatically determined in a matter of seconds.