Design with artificial intelligence metasurface based structure

Metamaterials can provide access to materials that cannot be formed naturally, with new optical properties such as negative or zero refractive index, invisibility, and shielding. The development of new optical metamaterials was progressing slowly, when the solution of Maxwell's equations required to provide these properties to metamaterials and the dependence of process optimization on the trial-and-error system were combined with the limitations of human factor capabilities. Graphic cards, which have developed under the guidance of the crypto money universe in the last ten years, have developed the artificial intelligence and deep learning universe. This development has helped to accelerate the algorithms developed with finite element modeling, finite difference time domain methods with deep learning and artificial intelligence applications, and accelerate the process of discovering new metamaterials. By using the target priority algorithm for high focusing efficiency wideband, low index lens design, 10,000 iterations were made in the 3.2um-16um design area, and the achromatic lens at 1300-1700nm wavelengths was achieved with the Lumerical program with 80% efficiency and 0.7 aperture. In order to eliminate the impossibility of production of the nanostructure, Bruggeman application was performed, again with 80% efficiency and 0.7 aperture. The simulation data was produced at 13,300 times magnification and tested, and 65% efficiency was achieved at 0.71 Transverse Electric and 0.69 Transverse Magnetic polarization. Again, high index lens design was made with target priority algorithm and designed with 1.86 aperture value at 1.2um focal length in 0.8um-9.6um design area and produced and tested with only 4% efficiency difference from design parameters. For a beam splitter to be designed using TiO2 cylinders on a SiO2 surface, 13767 simulations were created in a long period of one week with the Lumerical program. By multiplying 11014 of these data with deep learning algorithms, 6.2 million data that will take 60 years in the Lumerical program were created in 1 hour, and the dimensions of the cylindrical structures were determined and the beam splitter was designed.