Spatiotemporal forecasting of the tokamak edge instability using neural networks

Artificial Intelligence techniques have been increasingly adopted by the plasma and fusion science to address problems like plasma reconstruction, surrogate modeling, and tokamak/stellarator optimization. A key focus in sustained fusion research is the prediction and mitigation of Edge-Localized-Modes (ELM), instabilities that occur in short, periodic bursts and can cause erosion to the tokomak vessel wall. Recent research has demonstrated the power of neural networks in approximating continuous functions. In this work, we aim to create a high-fidelity sample generator and build a spatiotemporal forecasting model that can predict the onset of ELMs and their evolution at early stages by leveraging recent advances in generative modeling and Fourier neural operators. Conditional generative models, such as sequential Variational Autoencoders (Seq-VAEs), enable probabilistic autoregressive forecasting when combined with pretrained transformer foundation models. Adaptive Fourier Neural Operators (FNOs) offer robust forecasting potential for frequency-domain modeling of plasma. We apply these methods to predict ELM evolution using the beam emission spectroscopy system (BES) data and other diagnostics in DIII-D tokamak.