This study presents a machine learning model for predicting plasma disruptions using the KSTAR database. The model employs a long short-term memory (LSTM) network to capture temporal patterns in zero-dimensional plasma signals. A total of 668 disruption shots and 113 non-disruption shots from the 2019 to 2022 carbon divertor campaigns were used, divided into training, validation, and test sets. The architecture combines a multi-input LSTM and a fully connected neural network, using 30 features sampled over a 1 s window. The model achieved an AUC of 0.88 for individual samples and an F1 score of 0.91 in shot-by-shot evaluation, with over 90% accuracy for both disruption and non-disruption shots. Additional analysis using permutation importance and t-SNE visualization identified key features and confirmed the model’s interpretability. With an inference time of ∼3.1 ms per sample, the model shows strong potential for real-time application in plasma control systems.
This paper presents a machine learning model that uses long short-term memory (LSTM) networks to predict plasma disruptions in the KSTAR tokamak. The model analyzes zero-dimensional plasma signals and achieves high accuracy in both disruption and non-disruption predictions, making it a promising tool for real-time plasma control.