FAST DESIGN OF ELASTIC METASURFACE BASED ON MACHINE LEARNING

Elastic metasurfaces have attracted enormous attention from physical and engineering community due to their fantastic wavefront manipulation abilities. Traditionally, metasurfaces are designed by employing some specific mechanisms, such as stiffness tuning by cutting grooves, mass changing through adding surface pillars, resonators to control the dispersion and so on. But this design method depends highly on parameter-sweeping calculations and the manually analysis of the collected simulation data. Therefore, it is difficult to deal with high-dimensional problems. Recent fast development of machine learning technique (belonging to artificial intelligence (AI) techniques) provides another powerful tool for design of elastic metasurfaces, mainly owing to its excellent ability in building nonlinear mapping relation between high-dimensional input data and output data. Thus, we use the machine learning to train a network to learn the complex relation between the geometrical parameters of the elastic metasurface unit and its dynamic properties. After well training based on a big dataset collected from FEM simulations, this network can play the role of a surrogate model in the inverse design of elastic metasurface. Since the FEM simulations in inverse design is replaced by the surrogate model, the simulation is light and fast, making real-time design of AM possible. This novel design method can conveniently extend to design other metasurfaces for manipulation of optical, acoustical or elastic waves.