Real-Time Topology Optimization in 3D via Deep Transfer Learning

The published literature on topology optimization has exploded over the last two decades to include methods that use shape and topological derivatives or evolutionary algorithms formulated various geometric representations parametrizations. One of key challenges all these is massive computational cost associated with 3D problems. We introduce a transfer learning method based convolutional neural network (1) can handle high-resolution design domains shapes topologies; (2) supports real-time space explorations as domain boundary conditions change; (3) requires much smaller set examples for improvement in new task compared traditional deep networks; (4) multiple orders magnitude more efficient than established gradient-based methods, such SIMP. provide numerous 2D showcase effectiveness accuracy our proposed approach, including are unseen source network, well generalization capabilities learning-based approach. Our experiments achieved an average binary around 95% at prediction rates. These properties, turn, suggest transfer-learning may serve first practical underlying framework exploration optimization. • A approach CNNs. new, TO learning. Strong ability. Much training other approaches TO.