Detektion av defekter i nanostrukturer med maskininlärning - En lösning med Autoencoders och Unsupervised Learning

Abstract

This study aims to investigate the possibilities of using Machine Learning for detecting fabrication defects in nanostructures. The structures investigated are microwave circuits produced at the Department of microtechnology and nanoscience (MC2) at Chalmers University of Technology. Firstly, a particularly defective circuit was investigated with Logistic Regression, Dense Neural Networks, Convolutional Neural Networks as well a Transfer Learning based method with ResNetV2 and Principal Component Analysis. The distribution of defective and non-defective circuits was large and balanced enough to achieve 100 % correct classification of the sections with almost all models. Two more realistically defective circuits were further investigated, where the defective sections were widely underrepresented. However, a Convolutional Autoencoder (CAE), trained with either Supervised or Unsupervised Learning, was largely successful in separating defective sections from non-defective ones with a clear boundary based on the reconstruction error provided by the CAE. Furthermore, the CAE was in many cases able to locate the exact positions of defects by marking the areas of maximum reconstruction error, and also flag defective sections that previously was unsuccessful with manual inspection. The circuit sections used for the CAE were automatically sampled from larger images of circuits. After being sampled from the larger images of circuits, the sections were only downsampled and not preprocessed in any other way. The success of the unsupervised approach is the main achievement of the study, as all that is needed to train the model is completely uninvestigated images. The approach is estimated to save several days of manually inspecting whole circuits for defects.