Development of a Multimodal Network for Detection of Internal Reconnection Events in LTX-β

Our project aims to develop a machine learning-based classifier to automatically detect Internal Reconnection Events (IREs) in spherical tokamaks, like Lithium Tokamak eXperiment-β (LTX- β), a low-aspect-ratio tokamak capable of operating with walls coated in solid or liquid lithium. Using LTX- β data, we build a multimodal network that examines multi-dimensional relationships among both image and tabular data.

To train the model, I manually review and label 7,265 events from 2,965 shots, covering three years of diagnostic data from LTX- β. Each event includes several measurements like electron density, Mirnov coil data, etc.

Our initial model, a multilayer perceptron, achieves an accuracy of 89.9%. However, it relies heavily on manual preprocessing. To support automation, our current multimodal architecture incorporates a convolutional neural network that analyzes image data. The multimodal network processes both image and tabular data in parallel and concatenates the results, identifying IREs with 93.8% accuracy.

By automating IRE detection, this tool can assist operators in monitoring fusion performance and developing strategies to reduce confinement losses. Our next steps are to compare model architectures and ultimately enable real-time IRE prediction and detection.