Machine-Learning-Enabled Full-State Reconstruction of Fusion Plasmas from Minimal Sensor Measurements

Abstract

Plasma in nuclear fusion reactors is only partially observable: diagnostics are constrained by limited access, cost, and the harsh plasma environment, while high-fidelity simulations remain prohibitively costly at reactor-relevant scales to address the observability gap. This paper presents an ML model for reconstructing full-domain plasma states from a small number of accessible measurements. The model combines temporal encoding of sparse sensor histories with spatial decoding into complete plasma-field maps. Demonstrations using high-fidelity kinetic simulation data show that multiple coupled plasma quantities can be reconstructed from only a few density sensors, with robustness to sparse and randomly located probes. The approach provides a route toward diagnostic augmentation, real-time state estimation, and data-driven digital twins for fusion-relevant plasmas.