The Nano Foundries and Fine Analysis (NFFA) Europe project has been conceived with the objective of creating a platform to promote multidisciplinary research at the nanoscales connecting several specialised European institutions and laboratory facilities. Within the project particular effort has been focused on the creation of an Information and Data management Repository Platform (IDRP) to collect and maintain diversified data resources, with particular focus on meeting open and FAIR best practices.
The analysis of images collected via scanning electron microscope (SEM)has been one of the main case studies. During the last five years, continuous collaboration with the nanoscientists at the CNR-IOM facilities led to a careful analysis of the metadata of SEM images, and to the creation of several publicly available datasets of images humanly classified by content into 10 categories. Training the weights of a convolutional neural network on the labelled images enabled to create a service able to predict the belonging of newly acquired images to one of the 10 NFFA categories.
The classification into the 10 NFFA categories provides an extremely valuable proof of concept of how modern deep learning techniques can serve as a tool for supporting and stimulating scientific discoveries. Nonetheless, the classification is too coarse and cannot possibly cover the entire spectrum of content of the experiments collected at the various facilities. This motivates the study of state-of-the art techniques that allow at the same time to refine the NFFA classification while adapting to the continuous data acquisition.
This work aims at a combined investigation of supervised and unsupervised deep leaning techniques that have the potential of meeting these requirements.