Topology optimization is a tool to design lightweight structures with maximum performance. However, their scope of application is limited due to the high computational cost associated with large-scale problems. This work presents new approaches that improve the efficiency of existing methods by combining conventional acceleration methods with neural networks. Based on the standard solid isotropic material with penalization method, a higher-order multi-resolution scheme is employed, which utilizes higher-order shape functions and different resolutions for the finite element mesh and the density mesh. A parametric model order reduction method is proposed to condense the internal modes out of the stiffness matrix. Additionally, neural networks are used to reparameterize the density field in the optimization process. In an alternative approach, a UNet++ is trained to predict near-optimal density distributions based on local stress and strain fields. All methods are applied to the Messerschmitt-Bolkow-Blohm beam, where the results are analyzed regarding their solution quality and efficiency. The use of convolutional neural networks for the density parameterization leads to improved designs and the total computation time is reduced by a factor of 2.5. Similarly, designs with lower compliance and faster convergence are obtained by integrating predictions from the UNet++ into the optimization process.     «

Topology optimization is a tool to design lightweight structures with maximum performance. However, their scope of application is limited due to the high computational cost associated with large-scale problems. This work presents new approaches that improve the efficiency of existing methods by combining conventional acceleration methods with neural networks. Based on the standard solid isotropic material with penalization method, a higher-order multi-resolution scheme is employed, which utilize...     »