Description
Intermittent turbulent structures, known as blobs, play a central role in cross‑field transport at the tokamak boundary, where they dominate particle and heat flux into the scrape‑off layer (SOL). Their formation and outward convection are well established in both experiments and simulations [1-3]. Characterizing these structures, however, often relies on manual or threshold‑based methods that do not scale to modern high‑resolution datasets. Here we show that unsupervised machine learning, specifically K‑means clustering applied to TOKAM2D turbulence simulations, can automatically identify coherent density structures without the need for labels. Using only density and potential fluctuations, we find that the learned clusters naturally separate background plasma, blob wakes, steep-gradient regions, and blob cores. Importantly, potential‑enstrophy contours align with ML‑derived boundaries, indicating that the algorithm isolates physically meaningful structures. These clusters further reproduce the signatures of SOL intermittency: blobs occupy a substantial fraction of the turbulent region, carry a disproportionate share of density fluctuations, and produce net outward flux. Our results show that unsupervised ML provides an objective, scalable, and physics‑consistent tool for analyzing coherent structures in turbulence simulations, complementing recent machine learning advances in blob detection and tracking [4,5] and enabling automated, event‑resolved turbulence analysis in fusion‑relevant regimes.