Artificial Neural Network for Diffraction Based Overlay Measurement

Diffraction-based overlay (DBO) accuracy is critical to the intelligent nanolithography process control for producing advanced semiconductor fabrication nodes. Optical gratings located on various layers are commonly used as the targets for the detection of the overlay displacement offset in DBO measurement. The asymmetry in intensity between the 1st and -1st beams diffracted by the targets is used for the prediction of grating displacement offset. This paper describes the effect of grating targets with sidewall angles (SWAs) on asymmetry in intensity and proposes an artificial neural network (ANN) method for enhancing the accuracy of grating displacement offset prediction. Grating targets with a 1:3 line-to-pitch ratio and SWA profiles varying from 86° to 90° were employed in this paper. The asymmetry in the intensity of the designed targets was computed for incident beams with transverse-electric and transverse-magnetic polarization at visible wavelength. An ANN feed-forward model was developed for the displacement offset prediction. The ANN, the conventional linear model, and the regression models were evaluated using diffraction data calculated by a numerical electromagnetic solver. The mean square error and the mean of the residual indicated that using the ANN model and incident beams at wavelengths of 600, 650, and 750 nm is substantially more effective for prediction than the conventional linear model is.