A machine learning Normalizing Flow method for Uncertainty Quantification with applications to Runaway Electron Dissipation.

Understanding and quantifying uncertainties in physical models, numerical simulations, and experimental data is crucial. We propose a conditional pseudo-reversible normalizing flow (PR-NF) method to construct surrogate models of physical systems in the presence of noise, enabling efficient quantification of forward and inverse uncertainty propagation. The PR-NF model excels in determining conditional distributions for these processes. We apply the PR-NF model to study runaway electron dissipation with impurity injection in tokamaks. Developing a robust disruption mitigation system for ITER, and future machines, is essential to prevent reactor wall damage, yet the optimal impurity type and deposition method remain unresolved. We discuss how the spatiotemporal density profiles of injected impurities impact runaway electron dissipation. Also, to demonstrate the effectiveness of our approach, we present two validation tests and an application to runaway electron dissipation.