Low-noise CD-SEM images are required in order to obtain robust and reliable measurement results, especially for complex 2D patterns. However, standard practices to reduce CD-SEM noise during image acquisition (increasing number of frames, increasing beam current, etc.) also increase acquisition time and the probability of deteriorating the materials under inspection. This effect is getting further attention of the industry on the case of EUV (extreme ultraviolet) resist, being an electron sensitive material and presenting small thickness, requiring extra care in order to prevent damage. Although there are techniques that intrinsically improve the metrology robustness to noise (such as model-based contour extraction), denoising SEM images may prove useful and a complementary approach for further improve metrology quality in such challenging cases. In this work we propose an innovative solution using Deep Learning Neural Network (DLNN) for low frame image denoising, which performs the model calibration using low frame images only, from a reduced dataset. A specific data augmentation approach is used in order to limit the number of images needed for the training. The denoising performance of the algorithm was evaluated in terms of accuracy and precision over a synthetic dataset, not used during the training of the neural network. The results show an improvement both in precision (up to 50% on the extreme case, 5% on average) as well in accuracy (over 13% on average).
|