Speaker
Description
In this work, we present an integrated data-driven framework for pre-experimental L-H transition prediction. This methodology employs a Transformer-based generative artificial intelligence model to predict the spatiotemporal evolution of magnetic fluctuations, conditioned on planned actuator trajectories. These predicted signals are subsequently processed by a deep learning-based plasma confinement state classifier to identify the transition onset prior to the discharge.
In tokamaks, access to the high-confinement (H-mode) regime is favourable for economic viability of fusion energy. However, adequate modelling of the L-H transition dynamics remains to be fully developed. Empirical scaling laws, based on global parameters such as the toroidal magnetic field, electron density, and plasma surface area, are valuable for estimating the power threshold required for next-generation device design. Yet, these static formulations fail to capture the temporal evolution necessary for detailed scenario development. Conversely, recent data-driven approaches utilising real-time diagnostic signals have demonstrated success in temporal prediction suitable for feedback control. However, these methods depend on intra-shot data streams and thus cannot provide pre-experimental predictions of transition timing essential for experimental design and optimisation.
To address these limitations, this work presents a methodology for pre-experimental L-H transition onset prediction and its validation using KSTAR data. The generative artificial intelligence model synthesises the temporal evolution of Mirnov coil (MC) signals throughout the discharge, utilising planned actuator trajectories including fuelling, auxiliary heating, plasma current, and plasma shape. The MC signals are then processed by the Optimal Autoencoder-based State Identification System (OASIS), a deep learning framework, to classify L-mode and H-mode phases. Application to the KSTAR database demonstrates high accuracy in predicting the onset of L-H transition. These results suggest that the proposed approach offers a viable pathway towards data-driven tokamak experiment design.