Fusion reactor technologies are well suited to contribute in a secure and sustainable way to our future power needs. Numerical models can tell researchers about the conduct of the fusion plasma and provide useful insights into reactor design and operation effectiveness. However, a number of advanced models that are not quick enough to provide data on reactor design and operation are needed for the design of the large number of plasma interactions.
Training a Robot (Using Artificial Intelligence and Supercomputers) Is Just Amazing
Aaron Ho from the Nuclear Fusion Science and Technology Community in the Applied Physics department at the Eindhoven University of Technology has explored the use of machine learning methods for accelerating computational simulation of turbulent transport of core plasma. On 17 March, Ho defended his thesis.
The ultimate objective of fusion reactor research is to achieve an economically feasible net power gain. To this end, large and complex devices have been designed, however, as these fusion reactor devices become more complex, the first predictive approach to their operation becomes increasingly necessary. This eliminates operating inefficiencies and avoids serious damage to the system.
Challenges to Opportunities| Tech Companies Making Best Use
To simulate such a system, models must be used to capture all phenomena in a fusion device, are precise enough to use predictions to take confident design decisions and are fast enough to find workable solutions quickly.
Model Based on Neural Networks For Fusion
Aaron Ho used a model based on neural networks in his PhD research to create a model that satisfied these requirements. This method effectively helps a model to maintain both speed and accuracy while reducing data collection time. QuaLiKiz, a reduced-order turbulence model that predicts plasma transport quantities induced by microturbulence, was used to test the numerical method. In tokamak plasma systems, this phenomenon is the dominant transport mechanism. Unfortunately, in current tokamak plasma modelling, its measurement is also the limiting speed factor.
With QuaLiKiz evaluations as the training input, Ho was able to successfully train a neural network model with experimental data. The resulting neural network was then used to simulate the center of the plasma device using JINTRAC, a broader integrated modeling tool.
The simulation time was cut in half, from 217 hours to just two hours.
By replacing the original QuaLiKiz model with Ho’s neural network model and comparing the effects, the neural network’s performance was evaluated. In contrast to the original QuaLiKiz model, Ho’s model took into account more physics models, duplicated the results to within 10% of the original, and cut the simulation time in half, from 217 hours on 16 cores to two hours on a single core.
The model was then used in an optimization exercise involving the coupled device on a plasma ramp-up scenario as a proof-of-principle to assess its efficacy outside of the training results. This research contributed to a better understanding of the physics behind the experimental findings, as well as demonstrating the value of quick, accurate, and informative plasma models.
The Evolution of Marketing Funnels During the Harsh Times Of Pandemic
Finally, Ho recommends that the model be expanded to include advanced features such as controller and experimental design. He also proposes that the strategy be extended to other physics models, as the turbulent transport predictions were found to be no longer the limiting factor. This would increase the integrated model’s applicability in iterative implementations while also allowing for the validation efforts needed to bring it closer to being a genuinely predictive model.
The Entrepreneurs Diaries is now available on Telegram. Join our telegram channel to get instant updates from TED.
