Simulating the Universe with Machine Learning

Event Date:
2019-11-14T16:00:00
2019-11-14T17:00:00
Event Location:
Hennings 201
Speaker:
Shirley Ho (Flatiron/Princeton)
Related Upcoming Events:
Intended Audience:
Undergraduate
Local Contact:

Douglas Scott

Event Information:

The ever-increase need for accurate prediction for complex non-linear processes leads to large scale dynamical systems whose simulations and analysis make overwhelming and unmanageable demands on computational resources. The evolution of the Universe is one of these complex processes that the computational cost of the traditional full-order numerical simulations is extremely prohibitive. 

In this talk, we will talk about our attempts and sometimes successes in modeling the Universe's complex processes with machine learning instead of numerically solving the correct set of equations. We use state of the art deep learning methods to "learn" to "simulate" the interactions of planetary systems, the gravitational interaction of the entire Universe (with dark matter and dark energy) and the hydrodynamical interactions among the gas particles in the Universe.  We will also report and discuss a couple interesting cases in which our model generalizes far beyond our training set, and the implications therein. Are machines able to extrapolate far beyond what they are shown, just like humans who can figure out the physical laws that predict phenomena that are not yet observed? 

Add to Calendar 2019-11-14T16:00:00 2019-11-14T17:00:00 Simulating the Universe with Machine Learning Event Information: The ever-increase need for accurate prediction for complex non-linear processes leads to large scale dynamical systems whose simulations and analysis make overwhelming and unmanageable demands on computational resources. The evolution of the Universe is one of these complex processes that the computational cost of the traditional full-order numerical simulations is extremely prohibitive.  In this talk, we will talk about our attempts and sometimes successes in modeling the Universe's complex processes with machine learning instead of numerically solving the correct set of equations. We use state of the art deep learning methods to "learn" to "simulate" the interactions of planetary systems, the gravitational interaction of the entire Universe (with dark matter and dark energy) and the hydrodynamical interactions among the gas particles in the Universe.  We will also report and discuss a couple interesting cases in which our model generalizes far beyond our training set, and the implications therein. Are machines able to extrapolate far beyond what they are shown, just like humans who can figure out the physical laws that predict phenomena that are not yet observed?  Event Location: Hennings 201