Orateur
Martin Campos Pinto
Description
The delta-f method is a powerful tool to reduce statistical errors in particle simulations of kinetic problems. In its traditional form where it amounts of using an equilibrium state as a control variate, it is essentially limited to regimes where the distribution does not strongly deviate from this equilibrium. In general regimes, several methods have been proposed to extend the approach but the problem is still considered open by many experts in the field. In this talk I will review these approaches and present a new method where the control variate is evolved using neural networks, with promising numerical results both in low and
high dimensions.