Ju Sun leads team focused on developing novel models for materials discovery with $500K NSF grant
CSE DSI affiliate and Department of Computer Science & Engineering (CS&E) McKnight Land-Grant Professor Ju Sun is the principal investigator on ACED: Accelerating Materials Discovery by Learning with Physics-Informed Constraints, which recently received a $500K, 18-month grant from the National Science Foundation (NSF). The interdisciplinary team includes two more CSE DSI affiliates: Assistant Professor Chris Bartel from Chemical Engineering and Materials Sciences (CEMS), and Professor Zhaosong Lu from Industrial and Systems Engineering (ISyE). ACED will integrate physical laws into machine learning methods in order to advance material discovery.
“Our work is specifically looking at materials discovery,” Sun said. “There are large-scale data sets in the material domain that help train AI models to predict the properties of new materials, but these models are limited. We discovered that one of the main reasons these models fail is that they do not account for certain intrinsic physical laws governing the properties of materials. We proposed a novel way to mathematically model these physical laws into the training model training process, and from there we can significantly improve the performance.”
This project focuses on the foundational work in materials discovery. These new models will be built to predict the properties of new materials and propose uses for them. ACED will be flexible and have wide-ranging downstream applications, such as drug discovery, quantum computing, food production, and manufacturing.
“All three PIs on the team are from the University of Minnesota ,” Sun said. “Assistant Professor Chris Bartel is the domain scientist from Chemical Engineering and Material Sciences (CEMS). He provides the domain knowledge that is most crucial for us to model mathematically. We also have Professor Zhaosong Lu from Industrial and Systems Engineering (ISyE). He specializes in numerical methods and helps with the unique computational problems that come up.”
ACED is another example of the Knowledge-Guided Machine Learning (KGML) paradigm, a transformative framework for accelerating scientific discovery that was established at the University of Minnesota. Current artificial intelligence is largely data-driven, utilizing vast datasets to train models that make future predictions. However, the data-driven approach is fallible and often does not account for various constraints in the real world. KGML works to build basic physical laws and real-world constraints into models in order to get more accurate and usable results from AI.
“This is a frontier that many AI researchers are working on, because a data-driven approach can only get you so far. For our project, we need to have a way to predict the properties of new materials and propose uses for them. We are not applying these methods to existing materials, we are using AI to speed up the discovery process for the future.”