Randomized-SVD produces accurate characterizations of tokamak discharges from multiple data-streams

Modern magnetic fusion research involves high-resolution temporal and spatial diagnostics from multiple sensor arrays that generate large data streams well-suited to advanced scalable numerical linear algebra methods. This presentation introduces the first results of applying randomized numerical linear algebra (rNLA) techniques to the analysis of fusion data obtained from The Columbia University High Beta Tokamak-Extended Pulse (HBT-EP). rNLA techniques allow us to make use of distributing computing architectures and hence harness the power of GPU computing and massive parallelization for computing large matrix factorizations. Such factorizations provide the foundations of large-scale data analysis. To make use of rNLA methods we must trade off solution accuracy against computation time. In this work we show that a randomized singular value decomposition applied to data obtained from HBT-EP’s magnetic field sensors (poloidal, toroidal, and feedback arrays) produces results that are comparable to those obtained via exact direct (and thus not scalable) methods. The long term aims of this project are to integrate rNLA methods into pipelines involving highspeed video images of the plasma and push towards real-time control.