Metasurfaces have broad utility in spectroscopy, integrated optics, optical filtering, and holography. In this talk, I will discuss the development and utility of numerical tools for high performance design. For the first part, I will discuss the application of topology optimization to periodic and aperiodic metasurfaces. These methods are based on mathematical gradient descent, and they enable the discovery of new, freeform designs consisting of nonintuitive nanoscale patterns. For certain classes of devices, these metasurfaces have performance metrics (i.e., efficiencies and capabilities) that far exceed the performance of conventional metasurfaces based on phased arrays. Upon reverse-engineering these devices using coupled Bloch mode analysis, we find that the origins of high efficiency wavefront engineering is due to complex intramode and intermode coupling between the optical modes of the system.
For the second part, I will discuss the potential of machine learning to learn features in high performance devices and its use in automated device design. Machine learning is applicable to metasurface design and nanophotonics more broadly because there are clear relationships between geometric structure and optical response. I will show that these non-linear correlations between geometric structure and optical response can be learned in a neural network, and that a properly trained network can produce devices beyond the parameter space of the training data set. This reported research sets the foundation for the future of metasurface and nanophotonic design: one where computers and algorithms identify new design regimes of light-matter interaction unattainable by designs based on human intuition.