This talk is focused on a new approach for inverse design and knowledge discovery in large-scale metamaterials through integration of gradient-based (e.g., topological and shape) optimization and machine-learning techniques. It is shown that freeform metastructures can be represented by a few parameters through an aggressive dimensionality-reduction algorithm trained by the results of partially optimized structures through gradient-based techniques. The extensive reduction of dimensionality enables the design of optimal structures while facilitating better visualization of the effect of different design parameters.
Marzban et al. (Wed,) studied this question.