Science-constrained learning

The scientific promise of learning is to use data from measurements and simulations to train artificial models that overcome the complexity-accuracy trade-off faced by many (if not all) first principles models. Yet, these artificial models are rarely scientifically sound without massive amounts of data and handcrafted, ad hoc models, representations, and/or training objectives. Hence, before machine learning can fully realize its scientific promise, we must find principled ways to embed science into learning. This project tackles this challenge by formalizing and operationalizing science-constrained learning, an approach that uses constrained learning to explicitly incorporate scientific knowledge into machine learning.