 |
Center for Nonlinear Studies |
The last few decades of computational advances have transformed materials science. There is real potential for computational discovery of new materials that will enable disrupting technologies. The challenge lies in exploring the vast "space" of possible materials; in the current era, a direct approach seem impossible. But machine learning, successful in so many other domains, might make computational materials discovery an engineering tool. There are several hurdles; the biggest two are: data is sparse (and always will be) compared to conventional machine learning applications, and second, it is not clear how material should be represented mathematically in a machine learning environment. We have recently had a modicum of success in using machine learning to predict the properties of grain boundaries. An innovative representation was key. Furthermore, with our new representation provides a basis for physical interpretation of the machine learning predictions.
Host: Amy Larson