Physics-augmented machine learning techniques for modeling tokamak dynamics

For tokamak whole-device-modeling workflows, a combination of data-driven scalings and physics equations is used to estimate the plasma response to actuators. In prior work, a neural network to predict ne, Te, v, P, and q was trained on DIII-D experimental data from a broad selection of regimes. A database of TRANSP and ASTRA transport simulations predicting core Te and Ti for >100 DIII-D discharges was also generated, demonstrating results on par with data-driven scalings (an H98 confinement time scaling and a simple linear regression from edge to core values). These studies demonstrated the common intuition that data-driven models are more accurate than physics-based simulators, but fall behind in robustness and extrapolability. This talk presents methodologies for augmenting data-driven models with physics-based simulations, and demonstrates results for the task of predicting ne, Te, Ti, v, and q using a combination of neural networks and ASTRA. We demonstrate that while performance of models with the addition of information from simulations does not improve accuracy significantly within the training set distribution (DIII-D with plasma current < 1.3MA), there is a significant improvement for extrapolation beyond the training set (DIII-D > 1.4MA and AUG plasmas).