What is Surrogate Modeling?

Our friends in science and engineering are constantly developing advanced computational models. These easily become so complex that understanding how they work becomes a major challenge. The main idea of Surrogate Modeling is to develop a statistical model of the existing computational model, and subsequently to ask whatever scientific questions are of interest of the surrogate rather than asking them directly of the computational model as originally planned. This is especially important if the computational model takes a long time to execute. In practice, surrogate modeling proceeds by 1) using the computational model to create a dataset and then 2) fitting a statistical model to that dataset, 3) using that model to pick how to extend our dataset. This is an exciting area of research both in industry and the academy: most major tech companies, including Meta, Alphabet, and Amazon, have extensive expertise in this area, and the intersection of cool math, advanced computing, and impactful application makes for good papers at the intersection of math and CS.