Workshop: AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications
Authors: Leonardo Borgioli (University of Illinois at Chicago, Argonne National Laboratory); Neil James Getty (Argonne National Laboratory); Milos Zefran (University of Illinois Chicago); and Sandeep Madireddy (Argonne National Laboratory)
Abstract: Experimental tokamak diagnostics are multimodal, heterogeneous, and sparsely sampled across discharges, making reusable pretrained representations valuable for plasma modeling. Joint-Embedding Predictive Architectures (JEPAs) learn predictive latent representations without reconstructing every modality. We adopt TokaMark, a 14-task benchmark built from 11,573 MAST discharges, and introduce Tokamak JEPA (T-JEPA), a multimodal adaptation of V-JEPA~2.1 coupling a spatial poloidal magnetic-flux representation with six diagnostic encoders through bidirectional cross-attention. Across frozen, partial, and full adaptation, T-JEPA shows selective transfer, with frozen representations competitive for flux-map and plasma-current prediction. Our task-specific end-to-end models reduce NRMSE by 11.8--22.5\% relative to matched TokaMind FT-Base warm-starts on Groups 1--2, while normalization, actuator conditioning, and effective batch size affect accuracy and stability. Controlled studies show limited warm-start benefits and greater reliability when explicit coefficient targets are retained. Overall, task-specific optimization remains strongest for peak accuracy, while T-JEPA provides a basis for reusable multimodal representations across tokamak diagnostics.
Back to AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications Archive Listing
Back to Full Workshop Archive Listing