Conventional human-driven
methods face limitations in designing
complex functional metasurfaces. Inverse design is poised to empower
metasurface research by embracing fast-growing artificial intelligence.
In recent years, many research efforts have been devoted to enriching
inverse design principles and applications. In this perspective, we
review most commonly used metasurface inverse design strategies including
topology optimization, evolutionary optimization, and machine learning
techniques. We elaborate on their concepts and working principles,
as well as examples of their implementations. We also discuss two
emerging research trends: scaling up inverse design for large-area
aperiodic metasurfaces and end-to-end inverse design that co-optimizes
photonic hardware and post-image processing. Furthermore, recent demonstrations
of inverse-designed metasurfaces showing great potential in real-world
applications are highlighted. Finally, we discuss challenges in future
inverse design advancement, suggest potential research directions,
and outlook opportunities for implementing inverse design in nonlocal
metasurfaces, reconfigurable metasurfaces, quantum optics, and nonlinear
metasurfaces.