Overlay metrology is crucial to process control in manufacturing semiconductor devices. Diffraction-based overlay (DBO) is an effective overlay measurement approach because it exhibits multiple advantages. This study analyzed measurement errors caused by sidewalls in the bottom gratings of DBO targets. Accordingly, improvement was proposed using a neural network. First, rigorous coupled wave analysis was employed to calculate the pupil images generated by an overlay target. These images were then used as a data set. Next, two-directional twodimensional principal component analysis was used to reduce the dimension of features in these images. The features were then used to train a neural network and determine weighting coefficients in each network layer to create a DBO model. This study used virtual metrology to analyze 30 types of overlay targets and generated 18900 pupil images to create a data set. Each overlay target model was measured 10 times, and shot noise, dark noise, and quantization noise in the pupil images were accounted for. The simulation results revealed that when the dose level was 1000 mJ/s·cm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> , the overlay mean square error of the testing data was 0.40nm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> , indicating notable improvement in the measurement results of overlay targets with bottom grating sidewalls. Therefore, the proposed neural network-based DBO model can be applied to overlay targets with sidewalls and effectively improve the overlay accuracy.
Read full abstract