Likelihood-free inference using normalizing flows for experiment and simulation comparison

Modeling the scrape-off layer (SOL) in a tokamak device requires multi-physics processes such as turbulence, neutral transport, and material interactions. Ad-hoc anomalous transport coefficients are often employed in fluid code modeling, as turbulence is difficult to simulate, but require tweaking by hand in plasma codes to match observations. We introduce likelihood-free inference with normalizing flows to automate this process, determining the statistical significance of input parameters to match given observations and possibly revealing alternate parameter configurations that may otherwise be overlooked. Likelihood-free inference uses machine learning to approximate the posterior function p(θ|x) of a simulation’s input parameters, θ, given a set of outputs, x. Normalizing flows, which are constructed by neural networks and enable the mapping of a simple distribution to a far more complex one, are chosen to represent this desired distribution. The UEDGE 2D fluid simulation code is used to generate steady-state outer midplane and divertor measurements for a configuration given an input dataset of anomalous transport coefficients. The dataset is then used to train the normalizing flows, which are assessed using log probabilities.