Metamaterials design w/ machine learning
Where conventional design methods fail?
Structure-property relations of all existing (meta-)materials have primarily been explored in a forward fashion: given a microstructure, one extracts the effective properties by methods of homogenization. The inverse challenge – identifying a microstructural topology that meets the mechanical property requirements – has often been addressed by inefficient trial and error. Systematic approaches such as topology optimization and genetic algorithms are beneficial but also computationally expensive (relying on repeated sampling and computation of the effective properties), highly dependent on initial guesses, or may not be even compatible with complex design spaces (such as beam networks) or properties (e.g., curvature or buckling).
What can Machine Learning (ML) do for inverse design?
Machine-learning-based surrogate models, which can bypass expensive simulations and experiments, have been developed for applications including metamaterials, composites, and nanomaterials. However, such approaches require solving an optimization problem based on the (forward-only) surrogate model, which prevents an on-demand inverse design framework. In addition, while the forward problem, i.e., mapping from topological parameters to property space, is well-defined, the inverse problem is ill-posed (multiple topologies can have the same or similar effective properties).
To this end, we introduced a novel data-driven approach based on the integration of two neural networks for the forward- and inverse-problems that renders the inverse design challenge well-posed. This approach, in contrast to surrogate optimization methods, provides a computationally inexpensive two-way relationship between design parameters and mechanical properties. We applied this framework to a diverse range of metamaterial classes for tunable mechanical response (anisotropic stiffness only being the tip of the iceberg) as shown below.