Real-time topology optimization via learnable mappings (2024)

This research addresses the computational cost of traditional topology optimization. It introduces a machine learning approach that rapidly predicts optimal structures and stress fields by learning patterns from standard optimization results. Using a latent space representation and an autoencoder, the method generates accurate 2D and 3D designs, like the bridge visualized here, almost instantly, bypassing slow iterative processes.

The interactive visualization below demonstrates the 3D bridge example from the paper. Use the 'Parameter Controls' sliders to modify the bridge's support locations and applied forces. The predicted optimal structure updates instantly. Switch between viewing the material layout ('Topology'), Von Mises stress distribution ('Stress (VM)'), or tension/compression zones ('Tension/Comp') using the floating 'Select View Type' controls in the top-right corner. You can orbit the 3D model by clicking and dragging with your mouse. Please allow a few moments for the underlying machine learning models to load initially.