U.S. Department of Energy

Pacific Northwest National Laboratory

Materials Research Society Meeting welcomes Stevens

Abstract

Compressive sensing (CS) is a recently developed mathematical theory that allows for sampling rates lower than the Nyquist rate. In scanning transmission electron microscopy (STEM), CS has been shown to give electron dose reductions of up to 93%. The degree of undersampling depends on the materials science task to be carried out. Yet, there are many materials for which certain questions cannot be answered by standard STEM approaches. Some examples include: catalyst site determination, and protein structure identification. Utilizing CS will remedy many of these problems, but going a step further, and combining CS with machine learning (ML) will realize the full potential of CS. Moreover, ML models can be customized for specific materials science tasks rather than splitting reconstruction and statistical analysis into separate steps. In this work we primarily on dose minimization, but structure identification is a secondary concern. Within a Bayesian framework we apply active learning to a convolutional factor analysis model to determine the most informative sensing locations (e.g. pixels). Once the locations are found they are measured and the process repeats until a certain budget has been met. For example, if there was a budget of 20% of the pixels, we could measure a random 2% of the pixels, then use active learning in 9 more 2% batches. The convolutional model can help identify common structure within a specimen. This approach can further reduce dose in STEM and generally produces superior reconstructions relative to a completely random approach.

Stevens AJ, W Xie, Y Pu, L Carin, and ND Browning.  2015.  "Active Learning, Convolutional Factor Analysis, and Compressive Sensing for STEM Acquisition of Beam Sensitive Specimens."  Abstract submitted to 2015 Materials Research Society Fall Meeting, Boston, MA.  PNNL-SA-110946.

| Pacific Northwest National Laboratory