Hosted by: Deyu Lu

While machine learning holds promise in materials research it is currently limited to problems for which large datasets exist. I will discuss two applications where we have successfully used machine learning (ML). Phase identification in X-ray diffraction is particularly well suited for ML as simulated diffraction patterns can be used to generate training data, providing essentially infinite training data. I will show that convolutional neural networks, combined with statistical probability trees, can outperform traditional commercial software in identifying phases in single-phase and multi-phase diffraction patterns. In a second application, I will show how ML can be used to extract data and its associated meaning from millions of research papers in an automated way. Using this approach, we have generated the largest dataset on the synthesis of inorganic materials. Such datasets are critically needed to start using ML on difficult problems in materials science.