Taming The Complexity of Materials Degradation with Machine Learning

Limited stability and unacceptable degradation products are common reasons for otherwise promising materials to fail technological translation. The enduring state-of-the-art for establishing these properties essentially remains make-and-break testing, which is costly and provides information only at the end of the materials development process. The complexity of the reaction networks that govern degradation and the difficulty of data analysis pose tremendous obstacles to predictive approaches, however recent advances in automated reaction prediction and machine learning are increasingly making it possible to tractably describe and even predict degradation phenomena. In this talk I will highlight our group’s recent work developing methods for predicting reaction outcomes and how they have been applied to several different materials classes. The second half of the talk will discuss machine learning approaches to the closely related problem of identifying degradation products on the basis of typical spectral information sources.