The Attention-Guided Neural Network is designed to analyze periodic SEM images of SRAM. Autoencoder latent features layer is used to reconstruct the crops of the original image. By thresholding the ability of the autoencoder to reconstruct the original image, the defects and artifacts of the original image are automatically located. This approach could be used to detect visual anomalies.

An autoencoder-based approach is applied to CAFM images of the chip memory. Latent features visualizing highlights the regions, that could be of the interest to estimate the typical failures: missing contact, additional contact, not sufficient contact. Preliminary results are provided.