TokaMind: A Multi-Modal Transformer Foundation Model for Tokamak Plasma Dynamics

Researchers have introduced TokaMind, an open-source foundation model designed for tokamak plasma dynamics. The model utilizes a Multi-Modal Transformer (MMT) architecture and was pre-trained on the MAST dataset, incorporating diverse diagnostic data.

TokaMind is capable of processing multiple data types, including time-series, 2D profiles, and videos, even with varying sampling rates and missing signals. It offers efficient task adaptation through selective loading and freezing of its components.

Evaluated on the TokaMark benchmark, TokaMind demonstrated superior performance across most reconstruction and forecasting tasks compared to existing baselines.