Multigrid Distributed Deep CNNs for Structural Topology Optimization

Jaydeep Rade, Aditya Balu, Ethan Herron, Anushrut Jignasu, Sergio Botelho, Santi Adavani, Soumik Sarkar, Baskar Ganapathysubramanian, Adarsh Krishnamurthy

Keywords: Topology optimization, Deep Learning, Multigrid, Distributed training

Abstract: Structural topology optimization with traditional approaches is compute-intensive, mainly due to multiple finite element analysis iterations required to evaluate the component's performance during the optimization process. This computation cost scales up when performed on 3D high-resolution geometries. Researchers have developed deep learning (DL) based approaches, but these methods were demonstrated mainly using low-resolution 3D geometries (with a typical resolution of 32 X 32 X 32). We propose a DL-based method trained with a convolutional neural network (CNN) on high-resolution 3D geometries 128 X 128 X 128. With the initial strain energy (objective function of structural topology optimization) and target volume fraction (% material to be preserved after optimization) as the only inputs to the CNN, we predict the final optimized topology while maintaining the volume fraction constraint. To train the CNN at a high resolution is again a computational challenge. Therefore, we propose multi-resolution CNN, where we train the network at a lower resolution and then transfer the learned network to continue training at a higher resolution. Further, we significantly speed up the training time by 4.77X using distributed deep learning framework on GPU clusters (PSC Bridges-2).