Learning to Optimize: Artificial Neural Networks Applied to Topology Optimization
Please login to view abstract download link
We present a comprehensive deep-learning neural network architecture applicable to the problem of topology optimization in continuum solid structures. The proposed methodology learns to solve the inverse problem, making use of the physical constraints of the problem, which allows the model to learn with less data and obtain better generalization properties in new cases. Our model operates based on a predefined domain to optimize a set of loading positions and boundary conditions, obtaining very promising results in high-resolution samples and demonstrating its proficiency. The core architecture is based on an autoencoder model and is utilized to obtain a low-resolution representation of the solution, while a subsequent architecture is employed to enhance the final result. We learned some well-known methods like SIMP and UPM and are able to propose new solutions without iterating in quasi-real-time.
