Power Device Structure Optimization by Sequential Approximation Optimization Method

When design optimization is performed quickly, an approximator instead of a simulator can effectively speed up the computation. However, a large amount of training data is required to make approximators with high accuracy. In contrast, even with a small amount of training data, integration neural networks that enable highly accurate approximation have been proposed. We consider using sequential approximate optimization methods, which assume that optimization is performed to reduce the training data further. We propose a new stopping criterion for additional learning and a new method for adding training data corresponding to a multimodal optimization space. The results applied to power devices in this report are effective.