The potential of deep learning in the finite element method

Abstract

The Finite Element Method (FEM) has been widely used to obtain reliable solutions in various engineering fields. It is necessary to promote a leap of FEM technology by incorporating emerging technologies into FEM. Deep learning, one of artificial intelligence technologies, may have the surprising potential to innovate FEM. In this presentation, we investigate potential of deep learning in FEM through our previous studies. We developed two types of finite elements called "deep learned finite elements (DLFE)" and "self-updated finite element (SUFE)". We compare the performance of the proposed elements with those of other existing quadrilateral elements through numerical examples. The DLFE and SUFE produce improved accuracy and computational efficiency. Deep learning has the meaningful potential for finite element technology. In this study, the proposed methods were only applied to develop 2D quadrilateral solid finite elements. We will enlarge the methods to various finite elements such as triangle 2D solid, plate, and shell finite elements.