Physics Guided Neural Networks for the Design and Understanding of Materials

Abstract:

Climate change demands faster material innovations in multiple domains to reduce the carbon emissions of various industrial processes, but it currently takes 10-20 years to develop a single material with conventional human-driven approaches due to the large amounts of trail and errors needed. Machine learning approaches enable the direct learning of structure-property relations of materials from data, which have the potential to accelerate the search of materials and understand new scientific knowledge. However, one central challenge in developing an end-to-end framework for learning the structure-property relation is the uniqueness of material structure as a data type. It has both a discrete and a continuous component, differing from existing data types like images, point clouds, and graphs.

In this talk, I will present a neural network architecture that enables the end-to-end learning for materials and encodes several known invariances inspired by physics. I show that incorporating such invariances significantly improve the predictive performance of the model, and it also enables the interpretation of the model to provide scientific insights that are consistent with human understanding. I also demonstrate applying the method to solve a materials design problem for lithium ion batteries. Beyond simple scalar properties, I extend the approach to learn the dynamics of atoms in materials from time series data. I show that key structures and dynamical processes can be directly learned from data without supervision, which provides understanding to complex material systems that are difficult to study with conventional approaches. Finally, I will present my vision to create a unified model for materials that learns the “common sense” in materials science.

Bio:

Tian Xie is a fifth-year PhD candidate in the department of materials science and engineering at Massachusetts Institute of Technology (MIT), advised by Prof. Jeffrey Grossman. He worked at Google X and DeepMind as research interns during his PhD, and he received his B.S. in chemistry from Peking University in 2015. His research lies in the intersection of machine learning and materials science, and he is motivated by the urgent need to accelerate materials design due to the pressure from climate change. His work has focused on the development of physics guided neural network architectures that encode fundamental symmetries of materials, with applications to both accelerate the discovery of new materials and understand scientific knowledge from data. In addition to methodology development, he worked closely with experimentalists to validate the materials and mechanisms identified by the learned models. His vision is to create a unified model for materials that learns the “common sense” in materials science.