Tokamak reactors must fueled by cryogenic pellets. In addition, there is much evidence of pellet fueling playing a role in improved plasma confinement regimes. Fast and accurate modelling of pellet scenarios with integrated models would help understand and optimize these regimes. In addition, this modelling capability would benefit answering control-oriented questions such as fueling, burn, and exhaust control in pellet scenarios where incidents such as missing pellets can occur. This PhD position focuses on building up such capabilities for the JINTRAC multiphysics tokamak simulation suite1 , and then tackling the underlying research question of understanding and predicting pellet-enhanced confinement regimes. While JINTRAC, using first-principle-driven transport models, has been shown to accurately model pellet fueled scenarios2, the optimization and control-oriented nature of this project demands fast modelling and thus the extension of neural network surrogate models implemented within JINTRAC3. Experimental validation of the modelling pipelines will be carried out on existing data of a target tokamak to be determined (possibly ASDEX-Upgrade). To enable the research, the following tools must be developed.
- A fast surrogate pellet ablation and fueling model, using neural networks, of the HPI2 model
- Fast surrogate turbulence model valid for target tokamak regimes
- Development of the QuaLiKiz neural network tailored to pellet scenarios on the target tokamak, using Active Learning techniques for data efficiency
- Collecting a profile database of the target tokamak for sampling inputs for the NN training.
- Coupling the new surrogate models for JINTRAC "flight simulator" fast modelling
- Large-scale validation of JINTRAC + fast surrogate models in target tokamak pellet regimes
- Scenario optimization of pellet enhanced scenarios