Principal Investigator: Francis Alexander, Brookhaven National Laboratory
ExaLearn, one of six Exascale Computing Project (ECP) co-design centers, focuses on developing scalable machine learning tools and methods for science applications, especially for other ECP research areas and the US Department of Energy’s (DOE) experimental and high-performance computing facilities. ExaLearn includes experts from eight multipurpose DOE national laboratories, embodying the joint decision-making and multidisciplinary research implied as a co-design center.
ExaLearn is examining how methods for unsupervised, semi-supervised, and self-supervised learning technologies can be used to build domain science knowledge from large-scale data processed on extreme-scale supercomputers. To assure broad use, these methods will provide solutions that are both interoperable and reproducible.
Already, ExaLearn is exploring how machine learning can be used to create realistic surrogate models to replace computationally expensive simulations (e.g., requiring substantial time or memory, such as those from large-scale simulations of the universe). Another target area involves using machine learning to find the optimal mix of protocols and computing parameters that can enhance experimental design.