Metamaterials are lab-fabricated materials specifically engineered so that a property not found in naturally occurring materials can be exploited. Metamaterials can be used to affect sound waves, electromagnetic radiation, light, and several other natural phenomena in ways that natural materials can not. The behavior of metamaterials does not only depend on the base materials used to create them, but it is tied to properties that are determined by the shape, geometry, size, orientation and arrangement of their surfaces.
Calculating the optical properties of a metamaterial from known particle geometries and designs can be computationally expensive, and trying to invert the problem to come up with a design that produces desired optical properties generally amounts to a nonlinear, one-to-many problem that is very difficult to solve. Many researchers have recently been turning to machine learning (ML) to speed up these calculations, but research is still at the very early stages in this field and there is no accepted best practice. Separate models are typically required for the direct and inverse design problems, complicating interpretability and inefficiently decoupling the physics that must be relearned in both cases.
A framework that solves the inverse design problem in the context of optical properties of metamaterials can be generalized to many other fields of physics and engineering. Optical properties can be thought of as resulting (in part) from scattering of electromagnetic waves. As acoustics is about sound waves, a framework that works well for light might also work reasonably well for acoustics without requiring significant modification (other than a new appropriate training dataset, of course). Even more fundamentally, all of quantum mechanics is predicated on the wave function, which again therefore requires controlling waves; so it is conceivable that designing quantum technologies (e.g. solid state devices like transistors or photovoltaics) could be aided by ML inverse design architectures adept to handling waves. Waves show up in many other branches of physics too (earthquakes are elastic waves, ocean waves, heat in non-metals is fundamentally phonons, which are atomic vibrational waves, etc.), which gives an idea of how limitless the application of this innovation could be.
In this project, we built a web predictor that leverages deep learning techniques to solve the inverse design problem. We started from the findings of the research work already done by the Berkeley Lab and improved upon that by optimizing the existing models to get better results, leveraging some new models that delivered improved performance. We also extended the models by generalizing them and adding support for different base substrates.