Colloquium: Guiding Scientific Simulators with Machine Learning

10 May 2021

Colloquium by Dr. David Greenberg, Helmholtz Centre Hereon Geesthacht

Monday, 10 May 2021 17:15, Zoom Meeting

Abstract: Simulations are a powerful tool for combining and exploring scientific insights, and their predictions generalize better to new scenarios than non-physical data-driven aproaches. However, the problem of assimilating noisy and incomplete observations of the simulated system to constrain parameters or initial conditions is challenging, since for many important simulators the data likelihood is intractable.

I will describe two recent machine learning-based approaches addresing this problem: Bayesian inference with normalizing flows and optimization with differentiable emulators. Both of these approaches use model simulations as training data, allowing the machine learning model to benefit from the scientfic insight used to design the simulator.

Zoom access information: https://uni-hamburg.zoom.us/j/93735174299?pwd=Ums2cU9EZFpuc0ZFZWFOeWY5eU55dz09

Further information concerning the speakers and the lectures can also be found here: https://www.inf.uni-hamburg.de/home/kolloquium/sose21/greenberg.html