Multi-diagnostic Time Series Generative model for Prediction of the Edge Localized Modes in Tokamak plasmas

Artificial intelligence is increasingly applied in fusion research for plasma reconstruction, surrogate modeling, and device optimization. A major goal in fusion research is predicting and mitigating Edge-Localized Modes (ELMs)—short bursts that damage tokamak walls. Recent research in multi-modal generative models has demonstrated the power of neural networks in coupling different modes of data (images, texts, audio data). Leveraging these models, in this work, we extend our previous study on forecasting ELM cycles based on Beam Emission Spectroscopy (BES) signals to develop a multi-diagnostic model for predicting the onset of ELMs and imputing missing diagnostic signals. We leverage recent advances in generative modeling, sequence-to-sequence modeling, and Fourier neural operators to propose architectures and training strategies that can learn to forecast short to long term dynamics of the noisy signals due to ELMs. We benchmark the developed model against a state-of-the-art foundation model using the DIII-D Beam Emission Spectroscopy (BES) data that captures the plasma fluctuations due to ELMs over an 8x8 spatial grid. Our models demonstrate high accuracy, outperforming the baselines, in predicting the rapid rise and relaxation of the signals due to ELMs within 30-80μsec. Additionally, we present the model's ability to generate imputed diagnostics from observed signals.