Probing and Inference of Density Pedestal Structure Through Machine Learning

We utilize finite element computational modeling of a diffusive-convective ansatz to construct a database of density pedestals from defined profiles of transport and fueling for the purposes of training an inferential machine learning model. As fusion reactor technology closes in on commercial use, fine optimization of reactor performance will prove crucial. One way to produce improved reactor performance is through increasing global density by operating with a density pedestal. The pedestal profile is determined by particle transport and fueling, with the exact contribution of each term not well understood^[1]. In order to probe the relative contributions of these terms, they must first be untangled and examined. This is accomplished by building off of previous analytic work done^[2] and constructing a computational solver for the radial continuity equation and diffusive convective ansatz utilizing finite element methods. This computational model is then used to build a database of pedestal profiles with corresponding fueling and transport profiles. This dataset is used to train a machine learning model with the goal of inferring transport and fueling terms from an evolving density profile, as well as predicting how modifying these terms affects the development of the plasma. Through this, we are able to glean a better understanding of the contributions of the fueling and transport terms of the limitations of previously used analytical methods and the impacts of idealized assumptions made previously.