Neural selection of the optimal optical signature for a rapid characterization of a submicrometer period grating

Abstract

The characterization of gratings with small period-to-wavelength ratios can be achieved by solving the inverse problem of the diffraction. The use of a neural network has shown several advantages: it is a non-destructive, non-local and non-invasive method. However, although the calculation of results is instantaneous, the neural characterizations already published require the measurement of many diffracted intensities and can so need a long measurement time. We present, in this paper, a neural selection process called heuristic variable selection. This method reduces the number of diffractive efficiencies allowing a correct reconstruction of the profile shape according to an expected accuracy. In the same way, the non-redundancy of the data composing the optical signature is ensured. We relate a 1-μm period grating etched in silicon which could be characterized with only six measurements when a trapezoidal profile shape is assumed.