Convolutional neural networks at the interface of physical and digital data

Electron and X-ray interactions with matter can be recorded as digital images, which are signal acquisition mechanisms often used to investigate materials microstructure. Recently, the ability to quickly acquire large datasets at high resolution has created new challenges in areas that rely upon image-based information. The proposed analysis schemes employ Convolutional Neural Networks as the core algorithm in the reconnaissance of expected events from data gathered in two regimes: experimentally and by simulation. At the interface of physical and digital datasets, we propose classification schemes that exploit complex geometrical structure from scientific images through different machine learning packages, such as MatConvNet and TensorFlow. Our results show correct classification rates over 90% considering thousands of samples from four image modalities: cryo-electron microscopy, X-ray diffraction, X-ray scattering and X-ray microtomography. Our main contributions are: (a) developing algorithms designed for data that stem from physical experiments; (b) building new software to constrain parameter space, particularly given new hardware; and (c) testing different CNN models for classification of scientific images.