Experimental data improve large-scale simulations and increase physics understanding

What is it about?

How can we combine expensive experimental data with large-scale physics simulations to better understand the data and improve the simulation? How can we assign meaningful uncertainties to our improved predictions of future experiments? This paper describes a novel approach that uses machine learning and Bayesian inference to build a statistical model on a combination of experiments and simulations of laser-driven nuclear fusion implosions. We use the model to choose between competing hypothesis for the physics at play in the experiments and to make predictions, with uncertainties, of the potential performance of new experiments

Featured Image

Photo by Joshua Sortino on Unsplash

Photo by Joshua Sortino on Unsplash

Why is it important?

Our inertial confinement fusion experiments are typical of challenges that face many 'big science' efforts: experiments are expensive, difficult to diagnose, and live in a large design space; simulations are highly complex, combine multiple physics models, and to some extent fail to describe experimental observations; multiple potential explanations exist for observed discrepancy ranging from unknown input conditions to incorrect or missing physics. The method we present in this paper offers a path towards understanding these kinds of datasets for the first time.

Perspectives