Abstract
Computational simulation is an important tool for predicting the behavior of physical systems. Many powerful simulation programs exist today. However, using these programs to reliably analyze a physical situation requires considerable human effort and expertise to set up a simulation, determine whether the output makes sense, and repeatedly run the simulation with different inputs until a satisfactory result is achieved. Automating this process is not only of considerable practical importance but also raises significant AI research issues in the areas of spatial reasoning and deep models of expert reasoning about physics and numerical analysis.