Data-driven Models and Machine Learning for the Physical Sciences

David Hogg

Center for Computational Astrophysics, Flatiron Institute
New York University

There is immense hype, and immense promise, in machine learning for physics and astronomy. I use the case of stellar astrophysics as an example area in which to explore these ideas, but my points will be general and apply to any physics area where there are substantial data sets and good but not perfect physical models. When the information in the data is good enough to consistently rule out (in a statistical goodness-of-fit sense) the physical models, can we benefit from the data quality in ways that deliver new insights about fundamental physics? One of the main themes is that we want to pick and choose the parts of machine learning we do and don't want to be using, because our objectives are very different from those of Amazon and Facebook. I'll put a lot of emphasis on generalizability and causal structure. (Oh and by the way, data-driven models currently produce more precise measurements of stellar properties and compositions than any physical models.)