AI Model Boosts Fusion Experiments By Mastering Plasma States In Real Time

Their findings were recently published in the journal Nuclear Fusion.

Fusion Energy and the Challenge of Control

Fusion energy holds the promise of providing clean and virtually limitless power. However, for future reactors, they must operate reliably and avoid dangerous phenomena such as disruptions—sudden, intense events that can damage the reactor, and precisely control the plasma's confinement state to sustain high performance.

A Unified Model for Dual Tasks

To address these challenges, the team created a single AI system that performs two coupled tasks.

The model simultaneously identifies operational modes (such as L-mode and H-mode) and detects edge-localized modes (ELMs), reducing inconsistencies and improving robustness. Instead of using separate models for each task, the researchers used a multi-task learning approach. This allowed the model to leverage shared representations between tasks and achieve more consistent classifications, particularly improving recognition of the challenging “ELM-free H-mode” state.

Training on Real-World Experimental Data

The system was trained on 226 EAST shots collected between 2016 and 2022, producing 78,328 manually labeled 20-millisecond samples classified as L-mode, ELMy H-mode, or ELM-free H-mode. To strengthen interpretability and reduce noise, the researchers used physics-informed features: for operational mode classification, scalars such as plasma density, current, magnetic field, elongation, power loss, and minor radius were derived from L–H threshold and confinement scaling laws; for ELM detection, sequence signals from Dα radiation and Mirnov coils captured transient instabilities.

High Accuracy and Real-Time Capability

On held-out test data, the system achieved 96.7% accuracy and generalized well to previously unseen plasma shots. Importantly, inference took only ~0.68 milliseconds per sample, far faster than a conventional biLSTM baseline (~5.4 ms/sample), supporting near-real-time use in experiments.

Toward Smarter Fusion Reactors

These AI tools not only support more accurate and timely identification of plasma states but also deepen our understanding of plasma behavior. The technologies developed in this work establish an important step toward intelligent control systems for next-generation fusion reactors, complementing other AI-driven approaches such as disruption prediction and active plasma control.