Synthetic data generation using generative adversarial network for tokamak plasma current quench experiments

Abstract

Deep learning models for identification and subsequent mitigation of tokamak plasma disruption have recently shown great promise for reliable predictions for machines other than the one on which it has been trained. The performance of such artificial intelligence (AI)/machine learning (ML) models strongly depends on the training data. Considering the sparse availability of universal high quality data underscores the requirement for synthetic data for the training of the AI/ML models. Synthetic data generation methods reported in the current literature have limitations in terms of quantity, diversity and preserving the temporal dynamics of the experimental seed data (SD). The article presents generative adversarial networks based procedure capable enough to generate unlimited device-independent temporal evolution of tokamak plasma current. The synthetic data improves with the employment of the classified SD while retaining the characteristics of the original data. The procedure offers a substantial volume of synthetic data with a very impressive diversity, thereby ensuring the requirements for successful AI/ML model training.