Abstract: In this talk, I will introduce a new machine-learning method to discover structure-property relationships in complex materials data. The distillation of structure-property is a process projecting the changes of material properties to a controlled variable represented by a structured descriptor. Adversarial Autoencoder is a generative model that can be used to create a virtual chemical space that is easier to explore. We achieved controlled projection by leveraging prior physical knowledge to constrain the latent space so that each dimension tracks a targeted structure descriptor. The utility of this approach is illustrated for X-ray absorption spectra (XAS) of 3d transition metal oxides in exploring how the shape of an XAS spectrum changes in response to the change of a structure descriptor. The new method recovers trends established in the literature and reveals new ones.
Bio: Dr. Xiaohui Qu is a Staff Scientist at CFN. He works in the Theory and Computation Group focusing on research, development, and implementation of data analytics techniques. Dr. Qu is a computational chemist by training, with strong accomplishments in specific areas of chemistry, energy storage, machine learning, and X-ray absorption spectroscopy. He combines materials science research with extensive experience in deep learning techniques and a data-centric approach.