Machine-agnostic ELM onset prediction using explainable AI zero-shot turbulence analysis

Scientific Artificial Intelligence applications often remain limited by device-specific training, creating barriers to cross-system generalization critical for future fusion reactors. We demonstrate that our neural network, trained solely on MHz-scale beam emission spectroscopy turbulence measurements from DIII-D, successfully forecasts Type-I edge localized mode onsets in KSTAR without device-specific retraining. Through explainable AI combining gradient-weighted class activation mapping with physics validation, we reveal that our network autonomously internalizes universal physics mechanisms governing Type-I ELM instabilities rather than memorizing device-specific patterns. Statistical analyses of dimensionally-reduced saliency features show identical triangular patterns between saliency representations, instability growth rates, and prediction probability across both tokamaks. This establishes zero-shot machine-agnostic generalization for future reactor operations. Furthermore, we present that AI-based spatial resolution enhancement overcomes inherent BES spatial limitations, extending diagnostic capabilities beyond physical instrument constraints.