The worldwide research effort on sustainable energy supply by magnetic confinement fusion requires extensive numerical modeling of magnetized plasma on different levels as well as processing a huge variety of data from measurement and simulation. Preserving the mathematical structure of underlying laws of physics is key to obtain consistent and reliable results and can dramatically increase computational efficiency. This is illustrated on novel numerical integration schemes for non-canonical guiding-center equations of motion that underlie modern drift- and gyro-kinetic plasma models.
Subsequently, physics-informed machine learning methods based on Gaussian process regression are introduced that provide fast data-driven emulators for such mechanical systems as well as for partial differential equations. In addition to retaining conservation laws, these methods offer a unique combination of model- and data-driven approaches to differential equations, and are applicable in various areas beyond plasma physics.
TU Graz | Institut für Theoretische Physik – Computational Physics
18. June 2021, 09:00 AM - 10:30