Fusion reactor technologies will meet the future power requirements of the world. They are safe and reliable. Researchers can use numerical models to provide valuable insights into reactor design and operation and information about the behavior of fusion plasma. To model the vast array of plasma interactions, however, it is necessary to use a variety of specialized models that cannot be used fast enough to give data on reactor design or operation.
Aaron Ho, Science and Technology of Nuclear Fusion group at Eindhoven Univ. of Technology has investigated the potential of machine learning to accelerate the numerical simulation of core plasma turbulent transportation. He presented his thesis at the Eindhoven University of Technology on March 17.
Research on fusion reactors aims to realize a net power increase in an economically feasible way. Large, complex devices are necessary to achieve this goal. However, a predict-first approach is vital as these devices become more difficult. This helps reduce operational inefficiencies and prevents severe damage to the device.
For such a simulation, models must capture all relevant phenomena in a Fusion Device. They should be accurate enough so that design decisions can be made with confidence and can be quickly found workable solutions.
Neural networks are used to build models.
Aaron Ho, a Ph.D. researcher, developed a model that met these criteria using a neural network. This allows models to be accurate and fast while still collecting data. QuaLiKiz is a reduced-order model of turbulence that predicts plasma transport quantities due to micro turbulence. The numerical approach was used. This is the predominant transport mechanism in Tokamak Plasma Devices. It is the speed limit factor in current tokamak plasma modeling, unfortunately.
Ho trained a neural network model using QuaLiKiz evaluations and experimental data as training input. To simulate the core of the plasma device, the neural network was later integrated into JINTRAC’s larger integrated modeling framework.
Simulation time has been reduced from 217 hours down to just two hours.
The neural network’s performance was evaluated using Ho’s model, replacing the QuaLiKiz original model, and comparing the results. Ho’s model could simulate simulations in a fraction of the time it took to run the original QuaLiKiz models. It also included additional physics models and duplicated the results within 10%.
To verify the model’s effectiveness beyond the training data, the model was then used in an optimization exercise with the coupled system on a plasma ramp-up scenario to prove of principle. This study revealed a deeper understanding of the physics behind experimental observations and demonstrated the benefits of plasma models that are fast, precise, and detailed.
He also suggests expanding the model for additional applications, such as experimental design or controller. He also means that the technique be extended to other physics models as turbulent transport predictions are no longer the limiting factor. This would allow the integration model to be used in iterative applications and enable validation efforts to help the model become more predictive.
