Investigation of Methods for Expanding the Application of Integration Neural Network and Their Appropriate Structures

Simulation-based optimization is increasingly used in the thermal design of power devices. However, using highly accurate simulations such as FEM often takes too much time and makes sufficient optimization difficult. Therefore, surrogate model construction, which reduces the simulation time, is attracting attention. To address this problem, we focused on integration neural networks (INNs), which can construct highly accurate surrogate models with a small amount of training data by integrating deductive and inductive models. However, existing INNs cannot be applied as deductive models; they are only models that can be expressed by multiple regression analysis. Therefore, we extended the functionality of INNs to a model that can multiply and divide polynomial equations. As a result, we were able to construct a model with a coefficient of determination of 0.98 for the FEM simulation model.