Standard scalar wave propagation techniques, such as Fourier optics, struggle with the multi-scale challenges inherent in the inverse design of optical metasurfaces. Conventional approaches often assume that the source and observation planes are axially aligned and share the same spatial size and discretization. This becomes problematic in the case of metasurface design where often the source and observation planes are on the order of the wavelength. Designing metasurfaces nanometer by nanometer for large-scale applications is computationally prohibitive. Current metasurface inverse design methods generally approximate amplitude and phase under the local phase approximation. However, this is insufficient when considering the intricate interactions among nearest neighbors in a metasurface. A more comprehensive understanding requires the consideration of the complex near electric field, which holds richer information about the metasurface’s physics. Yet, computing the resultant complex field at a far distance from the metasurface is both essential for inverse design and challenging. This work presents an evaluation of three computational approaches, i.e., padded field propagation, shifted field propagation, and propagation by the chirp-z-transform, for the explicit purpose of metasurface inverse design. These techniques are implemented in the Pytorch Lightning deep learning framework facilitating optimization using the backpropagation algorithm.
|