Upscaling optimal topology multimaterials structures using Deep Neural Networks

Anirudh Kanthamraju (1) , Edouard Duriez (2) , Kai James (3) , Joseph Morlier (2)

1 ISAE-SUPAERO - Institut Supérieur de l'Aéronautique et de l'Espace
2 ICA - Institut Clément Ader
3 UIUC - University of Illinois at Urbana-Champaign [Urbana]

The problem of Topology Optimization aims to solve the question of the optimal material distribution subjected to known boundary and load conditions subject to a target volume fraction. In this study, we present a machine learning framework to tackle the problem of Multi Material Topology upscaling, i.e., the prediction of a higher resolution topology with just the low- resolution input. A Convolutional Deep Neural network was trained with a data set generated from an iterative code found in the existing literature. The network architecture implemented in this study is a modified version of SRGAN which has proven capabilities in upscaling complex real-world images. In this study, the perceptual loss function was used as the loss function in order to not penalize the network for its predictions that are off by a couple of pixels while simultaneously rewarding the network for its outputs that yield accurate compliance. The paper aims to present a novel approach to the problem of Topology Upscaling of 2x by mapping local features of the low-resolution topology to their higher resolution counterparts.

Mots clés: Multi-Material Topology Optimization; Deep Neural Networks; Image Super Resolution; Physics Informed Neural Networks

Domaines: Mécanique des structures [physics.class-ph]

Anirudh Kanthamraju, Edouard Duriez, Kai James, Joseph Morlier. Upscaling optimal topology multimaterials structures using Deep Neural Networks. CSMA 2022 15ème Colloque National en Calcul des Structures, May 2022, Presqu’île de Giens (Var), France. pp.0. ⟨hal-03693236⟩