Artificial neural networks application for stress smoothing in hexaedrons

In this paper it is presented the use of artificial neural networks to improve the tension fields obtained from the finite element discretization method. It was significantly reduced the time needed to reach solutions, with accuracy similar to the areas smoothing tensions methods: Superconvergent Patch Recovered (SPR) and Recovery by Equilibrium Patches (REP) improved. It is solved two cases that show the comparative advantages in terms of time spent by the neural network and the techniques described above for making improvements in the original solution: Artificial Neural Networks used only 7% and 70% respectively of the original time spent by the smoothing technique in such cases. As bigger is the magnitude of the problem, the greater the difference in the time required for the solutions, being better the neural network. Data used for this study come from cases of different features: with a smooth solution, a thick wall sphere exposed to inner pressure and one with singularities, a plate loaded with a lateral crack.