Abstract

A rigorous coupled wave model is used to generate simulated laser scatter patterns from diffraction gratings etched to various depths into a silicon wafer. The multivariate statistical analysis techniques of inverse least squares, classical least squares, principal component regression, and partial least squares are calibrated by using the simulated scatter data to predict the grating height. The same data set is also used to train a multilayer perceptron implementation of a neural network to similarly predict the grating height. Cross validation is used to evaluate the performance of each method. After the calibration, training, and validation phases, measured scatter patterns produced by structures are analyzed. The analysis not only addresses the inverse scatter problem in several ways but also indicates the preferred technique.

© 1992 Optical Society of America