How AI Predicts Tokamak Plasma Dynamics with Tiny Datasets

Background

Tokamak devices aim to achieve practical nuclear fusion by confining ultra‑hot plasma (temperature > 10⁸ °C, flow speed up to 100 km/s) with strong magnetic fields. During the discharge ramp‑down phase the plasma is highly sensitive to small control errors, which can trigger disruptive events.

Scientific Machine Learning Approach

A research team applied scientific machine learning (SciML) to combine first‑principles physics with experimental data. They built a Neural State‑Space Model (NSSM) that embeds neural networks into a zero‑dimensional physics backbone describing energy and particle balance, allowing the model to learn hard‑to‑model effects such as confinement time and radiation losses directly from data.

Dataset

The model was trained on 442 recent TCV discharge experiments, split into 311 training shots (only five in the high‑performance regime) and 131 validation shots. Despite the small size, the dataset is sufficient for the model to learn accurate plasma dynamics.

Model Architecture and Training

The NSSM defines a dynamics function f_θ and an observation function O_θ. Forward simulation generates predicted trajectories, which are compared to measured data to compute a loss. Automatic‑differentiation libraries (diffrax and JAX) provide adjoint sensitivities for efficient gradient‑based optimization of the neural parameters.

Key Experimental Results

Robustness validation : In discharge #81751 a tiny gap error amplified vertical instability, causing a sudden plasma shift. By injecting a stochastic gap‑error distribution into a reinforcement‑learning environment, retrained trajectories (e.g., discharge #82875) remained stable under similar errors, demonstrating learned robustness.

Predict‑first extrapolation : The current limit was raised from 140 kA to 170 kA and the entire discharge trajectory was generated solely from NSSM predictions before the experiment. Measured plasma parameters matched the predictions closely, and the discharge terminated safely, confirming the model’s ability to generalize to unseen operating conditions.

Performance

On a single NVIDIA A100 GPU the trained NSSM can simulate tens of thousands of ramp‑down trajectories per second, enabling rapid parallel exploration of control strategies.

Implications and Future Directions

The results show that AI‑augmented physics models can provide reliable, data‑efficient predictions for fusion plasma control, reducing the need for costly trial‑and‑error experiments. Ongoing work aims to integrate these models into real‑time control systems for future fusion power plants.

Related Works

Diag2Diag – a virtual diagnostic model that reconstructs missing plasma parameters from multi‑sensor data (arXiv:2405.05908v2).

FusionMAE – a large‑scale self‑supervised model that aligns 80+ diagnostic signals using a masked auto‑encoder architecture (arXiv:2509.12945).

References

Primary paper: https://www.nature.com/articles/s41467-025-63917-x (Nature Communications)

Additional source: https://news.mit.edu/2025/new-prediction-model-could-improve-reliability-fusion-power-plants-1007