Abstract: Our scientific understanding of the Universe has grown significantly in recent decades due to two fundamental thrusts: multifaceted theoretical efforts to predict observables in reality and large-scale efforts to collect data and measure properties with better and more versatile observational instruments. Rapid advances in resolution, precision, and scales are being made in both of these areas. Theory and simulation models can emulate increasingly complex physical processes; ongoing and future experiments produce massive, rich datasets. Prominent examples include the N-body simulation which evolves trillions of galaxies over billions of years. It is an effective approach to predicting structure formation of the Universe, though computationally expensive. In this talk, we show examples where deep neural networks can be used to predict structure formation of the Universe, ranging from gravity only simulations to prohibitively expensive hydrodynamic simulations. What is more surprising is that one of our learned models is able to accurately extrapolate beyond its training data, and predict gravity simulation of the Universe for significantly different latent cosmological parameters. Our study proves, for the first time, that deep learning is a practical and accurate alternative to approximate simulations of the gravitational structure formation of the Universe.
Machine Learn the Universe: Deep learning as the new simulator?
Shirley Ho (Center for Computational Astrophysics) // September 26, 2019
Abstract: Our scientific understanding of the Universe has grown significantly in recent decades due to two fundamental thrusts: multifaceted theoretical efforts to predict observables in reality and large-scale efforts to collect data and measure properties with better and more versatile observational instruments. Rapid advances in resolution, precision, and scales are being made in both of these areas. Theory and simulation models can emulate increasingly complex physical processes; ongoing and future experiments produce massive, rich datasets. Prominent examples include the N-body simulation which evolves trillions of galaxies over billions of years. It is an effective approach to predicting structure formation of the Universe, though computationally expensive. In this talk, we show examples where deep neural networks can be used to predict structure formation of the Universe, ranging from gravity only simulations to prohibitively expensive hydrodynamic simulations. What is more surprising is that one of our learned models is able to accurately extrapolate beyond its training data, and predict gravity simulation of the Universe for significantly different latent cosmological parameters. Our study proves, for the first time, that deep learning is a practical and accurate alternative to approximate simulations of the gravitational structure formation of the Universe.