Machine learning is variously viewed as both a revolutionary tool that will exponentially accelerate scientific discovery and as an overhyped bubble on the brink of bursting. This is due to its series of successful applications in the fast-paced tech industry, as well as the frequent misuse of the term “artificial intelligence” in its place. In general, machine learning is a collection of numerical methods for creating a predictive model from a large dataset when prior knowledge of the underlying model form is limited. These conditions have not, historically, been met in scientific research: data is almost always at a premium, while physical models are abundant. In recent years, there has been a simultaneous increase in the scale of data available to scientists and the accessibility of powerful open-source machine learning libraries, leading to the current spike in interest.<br/><br/>To find scientific problems amenable to a machine learning approach, one need only look to its successes in the tech industry, which, while numerous and varied, are most highly concentrated in the field of computer vision. Object detection and image classification are two areas that have particular relevance to materials science. In this talk, two datasets will be explored from a computer vision perspective. First, a collection of AFM images of fibrillar morphologies in polymer transistors will be used to demonstrate strategies for the extraction and analysis of fibers from images. This problem represents a case for which machine learning can augment a more traditional approach. Next, a database of nearly 20,000 AFM images will be used to demonstrate structural classification using deep neural nets, as well as new ways to search libraries of scientific imagery.

Low-hanging fruit for machine learning in materials science: Imaging

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.