Latent Space Mapping: Revolutionizing Predictive Models for Divertor Plasma Detachment Control

The inherent complexity of divertor plasma, characterized by multi-scale and multi-physics challenges, has historically limited high-fidelity applications to scientific research. Operational challenges in tokamak control and scenario development have necessitated reliance on simplified empirical methods. This work introduces a transformative machine-learning strategy that bridges this gap by developing rapid, precise surrogate models that encapsulate complex plasma and neutral physics. Utilizing latent space mapping, a method pioneered in inertial fusion and adapted for magnetic fusion, we efficiently represent complex divertor plasma states in a low-dimensional space, streamlining predictive model construction. We have developed specialized surrogate models through advanced 2D axisymmetric transport simulations and trained on about 100,000 UEDGE simulations from KSTAR tokamak equilibria. These models provide quasi-real-time predictions (within 100 ms) with exceptional accuracy (less than 20% relative error), forecasting crucial plasma parameters such as electron density, temperature, and heat loads at divertors and the detachment front locations — essential for effective plasma detachment control. Cross-validated with SOLPS-ITER simulations and calibrated against diverse tokamak campaign data, these models demonstrate robust efficacy. Their ability to precisely predict critical detachment phases marks a significant advancement in plasma physics, enhancing control strategies. The development of these models addresses challenges posed by limited diagnostics in future tokamak reactors and furnishes a robust toolkit for FFP designs. Their forthcoming integration into the Plasma Control System will set new benchmarks in real-time plasma management, marking a revolutionary step in fusion reactor operations.