Embracing Uncertainty in Large-Scale Computational Astrophysics.

A revolution is underway in astronomy resulting from massive astrophysical surveys providing a panchromatic view of the night sky. The next generation of surveys and the simulations used to calibrate them can produce in two nights what the previous generation produced over many years. This enormous image acquisition capability allows the telescope to revisit areas of the sky with sufficient frequency to expose dynamic features and transient events; e.g., asteroids whose trajectories may intersect Earth. At least three such surveys are planned; their collective output must be integrated and calibrated against computational simulations, prior surveys, and each other. Relational databases have been shown to be effective for astronomy at yesterday’s scale, but new access to the temporal dimension and increased intercomparison of multiple sources generate new sources of uncertainty that must be modeled explicitly in the database. Conventional relational database management systems are not cognizant of this uncertainty, requiring random variables to be prematurely and artificially collapsed prior to manpiulation. Previous results in probabilistic databases focus on discrete attribute values and are unproven at large scale. In this paper, we present concrete examples of probabilistic query processing from computational astrophysics, and use them to motivate new directions of research: continuous-valued attributes and queries involving complex aggregates over such attributes.