Learning to Design Lightweight Structures: Accelerating Form-Finding through Machine Learning and Differentiable Programming

Through evolutionary refinement, natural forms such as seashells have achieved exceptionally high stiffness-to-weight ratios. At the meso and architectural scale, designing such lightweight forms poses significant computational challenges. Their nonlinear mechanical behavior governs their optimal form, necessitating advanced design methods which are computationally expensive. In this talk, we present novel numerical methods that apply machine learning and differentiable programming to accurately accelerate the design of such forms across multiple scales. We showcase the practical value of our methods for the design of beam networks, masonry vaults, and gridhells; and validate our numerical results with physical experiments and demonstrators. Our work bridges the gap between computational mechanics and artificial intelligence, facilitating the design of mechanically efficient forms, bringing us closer to the material economy and elegance exemplified in natural forms.