Deeply Sub-Wavelength Non-Contact Optical Metrology of Sub-Wavelength Objects

We report single-shot measuring nanoscale dimensions of randomly positioned objects by an artificial intelligence examination of their optical diffraction patterns. The technique is based on training of a deep learning neural network with a large number of diffraction patterns of objects of a priori known dimensions. To demonstrate, we measured the width of 840 nanometric slits milled in an opaque metal film, using a conventional microscope and illuminated by a laser (λ=633 nm) while the light diffracted on the object was mapped at different distances from the object by a camera equipped with a high numerical aperture lens. Information on the slit width may be obtained from: a) overall intensity of light transmitting through the slit (larger slits transmit better), or b) analysis of fringes at the diffraction pattern. We tested both regimes using a standard deep learning protocol. The results of these experiments are presented in Fig. 1 as plots of the retrieved widths of 84 different slits that have never been seen by the neural network against their widths measured by scanning electron microscopy (SEM).