Shen, Haotian2019-08-292019-08-2920192019-08-29http://hdl.handle.net/1828/11078Proto planetary disks provide hints to how planets are formed. To understand them, it is necessary to run multiple simulations with various hyperparameters through trial-and-error. This procedure is typically time consuming as each simulation is expensive. In this project, we aim to solve this problem by learning a deep network that interpolate and extrapolate, given two simulation outcomes. We then iteratively apply this network to extrapolate simulations. Specifically, as proto planetary disks are circular, we make use of the log-polar representation of these disks and apply a circular 1D convolution on it. We empirically motivate our design choices via ablation study. Our experimental results show encouraging outcomes on approximating proto planetary simulations through extrapolation.en-USAvailable to the World Wide Webdeep networksimulationproto planetary disk1d convolutioncyclical paddingextrapolationinterpolationiterative extrapolationproject