Electrically tunable optical metasurfaces based on Liquid Crystal (LC) offer fast switching speed, low cost, and mature technological development, making it highly desirable for these applications. However, to date, all electrically tunable metasurfaces are designed at a single phase using physical intuition, without controlling the alternate phase and thus leading to limited switching efficiencies (~30 %) and small angular steering (15 degrees). Here, we use adjoint-based “inverse design” (equivalent to “backpropagation” in deep learning) to discover tunable metasurfaces with state-of-the-art efficiency (>80 %) and wide-angle steering (144 degrees). Inverse design can efficiently compute sensitivities with respect to arbitrarily many geometrical degrees of freedom, and thus is very effective for optimizing complicated photonic devices.
|