Skip to yearly menu bar Skip to main content


Poster
in
Workshop: AI for Science: Scaling in AI for Scientific Discovery

Fast-forward FARGO: Accelerating Protoplanetary Disk Simulations with Limited Data

Valentina Tardugno Poleo · David W. Hogg · Shirley Ho

Keywords: [ Astrophysics ] [ Planet Formation ] [ emulator ] [ CNN ]


Abstract: Hydrodynamic simulations of planets embedded in accretion disks around young stars are an essential tool in the study of planet formation and evolution. However, these simulations are expensive, and any quantitative comparisons between planet-formation theory and the observations of planets will require the execution of at least thousands of simulations. We present a U-net emulator built to accelerate the simulation package FARGO. Our emulator, Freesbee, was pre-trained with disk surface densities from 940 short, low-resolution FARGO simulations and fine-tuned using only 10 to 30 longer and higher-resolution simulations. The emulator takes as input the disk density and embedded planet position from the first few timesteps of a brief simulation and outputs the final state 5,000 dynamical times (years at 1 a.u.) later. The emulated disk densities have median fractional errors (relative to high-resolution FARGO) ranging from 0.9 to 6 percentwith a median value of 3 percent; they are over $10^6$ times faster to compute.

Chat is not available.