Developing Image-Based and Deep Learning Enhanced Diagnostics for Determining Chemical Compositions

Salt creep is a naturally occurring process that describes the dried pattern left after drying. This phenomenon is relevant to soft matter science and coupled with sparsely available image data, the main obstacles to reliable predictions and theoretical understanding are further complicated. In addition, this study explores the performance of images under various augmentations (zoom, rotation, image flip, zero-phase component analysis). The goals of this study are twofold: 1) to establish robust applications of artificial intelligence and machine learning methods in soft matter science and 2) through implementation, to provide capabilities to the deposit composition based on collected salt creep images. In this study, we exploited CNN-2D, ResNet-50, and GoogLeNet models to differentiate between four inorganic salts used in this study. These insights gained from this study have the potential to automate a process that is typically performed manually and allow for easy mitigation of issues that stem from the effects of salt creep.