Abstract

The realization of Fusion Pilot Plants requires control systems capable of continuous, stable operation in high-neutron-flux environments where diagnostic and actuator degradation is likely. To address these challenges, we detail a critical advance in resilient tokamak operation leveraging both interpretable artificial intelligence (AI) and hardware-failure-robust machine learning. First, we utilize a deep survival machine event prediction framework that reliably predicts tearing modes (TMs) more than 500ms before onset. We leverage an interpretable AI analysis framework to correlate plasma profiles directly to TM stability, uncovering which profile features are most important. Fine detailed AI interpretation techniques, including Shapley analysis, provide a direct connection between TM stability and the shape of Te, rotation, and q profiles. Building upon advanced scenario control, we introduce ECHO, a hardware-failure-robust machine-learning controller that dynamically coordinates multiple electron cyclotron heating (ECH) sources. If individual gyrotrons fail mid-shot, ECHO instantly adapts the duty cycles and steering of surviving units to sustain the target electron heating and current drive deposition profile. Utilizing ECHO and a real-time AI impurity model, we avoided confinement loss from core Argon accumulation by controlling ECH deposition by adjusting power for impurity shielding and core density pump-out while tracking mode locations. Furthermore, by actively controlling the edge and core deposition of ECCD to weaken sawtooth events and achieve robust access to RMP ELM suppressed plasmas.