Applying machine learning in scanning transmission electron microscopy (STEM)
This paper applies convolutional neural networks to simulated HAADF-STEM images generated from a simulation coded for this project. Due to the recent advances in scanning transmission electron microscopy (STEM) techniques there has been a significant increase in the size and quality of data sets produced. It is not uncommon to have thousands of STEM images produced and the analysis of these images is a time-consuming process.
A necessity for adequate, automated methods to extract information from these images is therefore relevant. Artiﬁcial intelligence and machine learning are being used increasingly in industry and research to automate processes and analyse large sets of data in a much more efﬁcient way than the human mind ever could. This could potentially cut down the time taken for characterising these images, leaving more time for important research and further study that could help the advancement of nanophysics.
However, current methods for optimizing neural networks involve trial and error with very little focus on understanding the role of each parameter on the accuracy of the model. New proposed neural networks are used to classify tetrahedrons with and without spherical voids. A new method using modeFRONTIER is proposed to replace the usual ﬁne-tuning of a neural network using the time-consuming trial and error approach. modeFRONTIER successfully maximises the accuracy and minimises the loss of the neural networks by tuning the dropout percentage and the batch size of the data inputted. The single-input network is optimised to achieve a ﬁnal accuracy of 99.2 ± 0.3% and a loss of 2.1±0.2% on a data set of 20,000 images when using modeFRONTIER; the accuracy does not increase beyond 93% when simply adjusting the parameters manually.
Therefore, using modeFRONTIER to vary the parameter values within the neural network has a clear and significant impact on its performance. Further work could involve optimising multiple parameters for different network structures.