Parameter Estimation of Machine Learning Models for Tokamak Pulse Simulation using Adaptive Methods

Sensitivity analysis and uncertainty quantification of machine learning approaches for surrogate modeling have enabled more efficient sampling of high-dimensional parameter spaces, where the combinatorial complexity of all possible conditions and the numerical instability of available physics codes prohibit uniform coverage. We adopt the Integrated Plasma Simulator (IPS) workflow to deploy massively parallel tokamak pulse simulation over selectively scanned parameter ranges in key physics modules of the core, edge, and scrape-off layer (CESOL). Recent success with active learning and adaptive sampling has advanced the development of surrogate models towards near real-time transport predictions. We show that data-driven and probabilistic approaches provide a unified framework for machine learning regression to reduce the sensitivity to high degrees of freedom and assess the uncertainty in outlier samples through identification of dominant parameter contributions in these models. These results will be presented for the Trapped Gyro-Landau Fluid (TGLF), Edge Pedestal Transport Barrier (EPED), and Scrape-off Layer Plasma Simulator (SOLPS) codes.