Machine learning combining simulation and experimental data for high-qmin scenario control development at DIII-D

We attempt to build on decades of experiments, theory, and simulations to more accurately predict and control systems. As a concrete use-case for such a model, we consider development of a model-predictive controller for achieving higher betan in DIII-D's high-qmin scenario. We use experimental data combined with simulations from the MHD stability code DCON (for tearing mode betan limit calculations), the transport codes ASTRA and TRANSP (for flux-driven profile evolution), and the fast-ion distribution code RABBIT (helpful for Alfven eigenmode analysis and predictions). We compare the accuracy and robustness of predictors and controllers trained on just simulation data, just on experimental data, and with both pieces of information.