New Neural Networks for Plasma Profile Prediction

As machine learning has become more accessible, it has become an adaptable, powerful tool that can reduce computational costs for modeling complex systems. Since modeling the plasma in a tokamak reactor is both impractical and unreliable, it is more advantageous to use previous reactor data and current diagnostics to predict how the plasma will evolve in the future. As the accuracy and speed of these models develop, they have the potential to become the core of tokamak reactor control systems. We sample different types of neural networks to predict plasma conditions inside tokamaks to find models that are better suited for active control of tokamaks. To analyze the utility of each type of model, we compare the computational costs of training, the accuracy of the models, and the speed for a new prediction to be made.