IFS Seminar

Abstract: The design of a fusion device requires the coupling of several multiscale simulations, such as ideal magnetohydrodynamic equilibria, neoclassical transport, turbulent transport, and profile evolution. Additionally, in order to optimize for a given design point, all these calculations should be performed multiple times to optimize several objective functions, such as confinement, stability, and performance, normally using gradient-based optimization. While autodifferentiation has recently emerged as one solution to find such gradients, its memory and computational time requirements often become impractical once it is applied to nonlinear systems that require hundreds or thousands of iterations. Here, we show how the idea of differentiating the solver, not the equation, has been successfully applied to several elements of the optimization chain to obtain fast and efficient solutions and their gradients. The new UWPlasma code database will be presented, including VMEX, GKX, DKX, SOLVAX, and NEOPAX.

 

Affiliation: University of Wisconsin-Madison

 

Short Bio: Rogerio Jorge has been an assistant professor at UW-Madison since January 2024. His group focuses on fast, differentiable simulations for plasma physics and fusion energy, as well as its applications to machine learning and reactor design. Previously, he held appointments at the University of Maryland, College Park, the Max-Planck Institute in Germany, and the University of Lisbon. He finished his PhD at Ecole Polytechnique Federale de Lausanne in 2019, where he worked in scrape-off layer turbulence and gyrokinetic theory.