Super-resolving 2D stress tensor field conserving equilibrium constraints using physics-informed U-Net

In a finite element analysis, using large number of grids is important to obtain accurate results, but resource-consuming task. Aiming real-time simulation and optimization, it desired fine grid analysis results within limited resource. This paper proposes super-resolution method that predicts stress tensor field in high-resolution from low-resolution contour plots by utilizing U-Net-based neural network which called PI-UNet. addition, the proposed model minimizes residual equilibrium constraints so outputs physically reasonable solution. The trained with FEM simple shapes, validated complicated realistic shape evaluate generalization capability.