Neural inverse design of self shaping composites

Self-shaping fiber-reinforced polymer (FRP) sheets offer a moldless method for generating complex 3D geometries by leveraging internal stresses. This innovative process enables sustainable fabrication of curved forms but presents a challenging inverse design problem: determining the fiber orientations that will produce a desired geometry. A data-driven approach is introduced for the inverse design of self-shaping FRP sheets, leveraging Convolutional Neural Networks (CNNs) trained on a large synthetic dataset generated from physically informed simulations. Two strategies are evaluated: (i) a direct prediction model that predicts fiber layouts from target surfaces, and (ii) an optimization method based on a trained surrogate model. The models are validated against physical prototypes and integrated into the Grasshopper design environment, facilitating accessible use within design workflows. It is demonstrated that the direct prediction model offers rapid, accurate predictions within the training distribution, while the optimization method generalizes better and is able to approximate freeform geometries. This work highlights the potential of machine learning to streamline moldless fabrication processes for complex architectural components.