Faster calculation of fusion reactors thanks to machine learning

Fusion reactor technologies are well positioned to contribute to our future power needs in a safe and sustainable way. Numerical models can provide researchers with information on the behavior of the fusion plasma, as well as valuable insight into the effectiveness of the design and operation of the reactor. However, to model the large number of plasma interactions, you need a number of specialized models that are not fast enough to provide data on the design and operation of the reactor. Aaron Ho of the Science and Technology of Nuclear Fusion Group in the Department of Applied Physics investigated the use of machine learning approaches to accelerate the numerical simulation of nuclear-turbulent plasma transport. Ho defends his thesis on March 17th.

The ultimate goal of fusion reactor research is to achieve a net power gain in an economically viable way. To achieve this goal, large complex devices have been built, but as these devices become more complex, it becomes increasingly important to follow a prediction-first approach regarding their operation. This reduces operating efficiency and protects the device from serious damage.

To simulate such a system, models are needed that can capture all the relevant phenomena in a fusion device, are accurate enough so that predictions can be used to make reliable design decisions, and are fast enough to find workable solutions quickly.

Model based on neural networks

For his Ph.D. research Aaron Ho developed a model to meet these criteria using a model based on neural networks. With this technique, a model can effectively maintain both speed and accuracy at the expense of data collection. The numerical approach was applied to a reduced order turbulence model, QuaLiKiz, which predicts plasma transport rates caused by microturbulence. This particular phenomenon is the dominant transport mechanism in tokamak plasma devices. Unfortunately, its calculation is also the limiting speed factor in the current tokamak plasma modeling.

Ho successfully trained a neural network model using QuaLiKiz evaluations while using experimental data as the training inputs. The resulting neural network was then linked to a larger integrated model framework, JINTRAC, to simulate the core of the plasma device.

Simulation time reduced from 217 hours to only two hours

The performance of the neural network was evaluated by replacing the original QuaLiKiz model by comparing Ho’s neural network model and the results. Compared to the original QuaLiKiz model, Ho’s model considered additional physics models, duplicated the results to a 10% accuracy, and reduced the simulation time from 217 hours on 16 core points to two hours on one core.

To test the effectiveness of the model outside the training data, the model was used in an optimization exercise using the coupled system on a plasma start-up scenario as proof of principle. This study provided a deeper understanding of the physics behind the experimental observations and highlighted the advantage of fast, accurate, and detailed plasma models.

Finally, Ho suggests that the model can be extended for further applications such as controller or experimental design. He also recommends extending the technique to other physics models, as it has been noted that turbulent transport forecasts are no longer the limiting factor. This will further enhance the applicability of the integrated model in iterative applications, enabling the required validation efforts to push its capabilities closer to an actual predictive model.

---

---