KI prognostiziert das "Ausbrechen" von Plasmen. MIT und andere Institute erreichen hochpräzise Vorhersagen der Plasmadynamik bei kleinen Stichproben mithilfe von maschinellem Lernen.

A research team led by the Massachusetts Institute of Technology (MIT) has utilized scientific machine learning to intelligently integrate physical laws with experimental data. They developed a neural state-space model that can predict the plasma dynamics during the ramp-down process of the Tokamak Configuration Variable (TCV) and potential instabilities with only a small amount of data.

If you're directly introduced to the "Tokamak device," you might find it unfamiliar. However, if you're told that the Tokamak device is one of the key technologies leading to the most ideal energy source - nuclear fusion energy, you might have an "Ah, so it's you" moment. Note that the "nuclear energy" here isn't the nuclear fission used in nuclear power plants, but nuclear fusion, which is more energetic, cleaner, safer, and produces almost no radioactive waste.

Nuclear fusion mimics the energy generation process inside the sun, releasing energy by fusing light nuclei (such as deuterium and tritium) at extremely high temperatures. To achieve this, we need to "create a small sun" on Earth. The Tokamak device contains plasma hotter than the sun's core in a toroidal vacuum chamber and uses a strong magnetic field to confine it, thus maintaining a stable fusion reaction.

However, while the ideal is promising, the reality is extremely "sensitive." For the Tokamak device, the current ramp-down at the end of the discharge is a "high-risk" phase. It deals with a plasma flow moving at a speed of up to 100 kilometers per second and with a temperature exceeding 100 million degrees Celsius. At this time, the plasma is in a strong transient state, and any tiny control error can trigger a destructive disturbance, damaging the device.

In this context, the research team led by MIT used scientific machine learning (SciML) to intelligently integrate physical laws with experimental data. They developed a neural state-space model (NSSM) that can predict the plasma dynamics during the ramp-down process of the Tokamak Configuration Variable (TCV) with only a small amount of data and potential instabilities, providing additional support for the safe control of the "artificial sun" shutdown.

The relevant research, titled "Learning plasma dynamics and robust rampdown trajectories with predict-first experiments at TCV," was published in Nature Communications.

Research Highlights:

* Proposed a neural state-space model (NSSM) that combines physical constraints with data-driven methods, enabling high-precision dynamic prediction and fast parallel simulation during the Tokamak discharge ramp-down phase.

* Conducted an extrapolation verification of "predict-first" in the TCV experiment. The closed-loop method of "predict first, then experiment" achieved a true data-driven control verification.

Paper Link: https://www.nature.com/articles/s41467-025-63917-x

More AI Frontier Papers: https://hyper.ai/papers

Dataset: Efficient Learning with Small Samples

The dataset used by the research team to train the model contains the latest 442 discharge experiment records of the TCV device. They divided 311 records for training (only 5 of which belong to the high-performance range) and 131 for validation. Can you sense how "mini" the dataset is?

With just this relatively small-scale data, the model learned to predict complex plasma dynamics and can simulate tens of thousands of ramp-down trajectories in parallel per second on a single A100 GPU, demonstrating its powerful learning and prediction capabilities.

Model Validation Metrics

The "Neural State-Space Model" with Physics as the Framework and Neural Networks as the Core

The core of the research is to build a model that can accurately predict the complex dynamics of the plasma during the shutdown phase. To this end, the research team designed a "neural state-space model" that integrates physics and data.

The framework of this model is a zero-dimensional physical equation, which mainly describes the energy balance and particle balance of the plasma. However, some key parameters (such as confinement time and radiation loss) are difficult to accurately model using first-principles. Therefore, the research team embedded "neural networks" in these core parts to enable the model to learn these hard-to-simulate physical effects from experimental data. It's like an autonomous driving car with a standard vehicle chassis, but its "driving experience" is trained from real road conditions.

Specifically, the model takes a series of controllable "actions" as inputs, such as the change rate of the plasma current and the neutral beam injection power. By solving this hybrid differential equation system composed of "physical equations + neural networks," the model can predict the future step by step.

Core Equation

The training process of the neural state-space model (NSSM) follows an efficient and automated procedure. The model is defined by the dynamics function fθ and the observation function Oθ. It performs forward simulation to generate prediction data and calculates the loss by comparing it with the experimental observations. With the automatic differentiation adjoint method of diffrax and JAX, the model parameters are optimized.

Description of the NSSM Training Method

Interesting and Inspiring Experiment Section

Among all the experiments, the two most inspiring results come from the robustness verification of the "sensitivity to control errors" and the "predict-first" extrapolation test. The former reveals a vulnerability during the ramp-down phase - when there is a small deviation in the high-field side gap, the growth rate of vertical instability may increase by an order of magnitude, triggering a vertical displacement event (VDE).

In a discharge numbered #81751, this phenomenon led to a sudden shift and termination of the plasma. Based on this, the research team introduced an uncertainty distribution of the gap error in the reinforcement learning (RL) environment, allowing the trajectory to actively adapt to the uncertainty during training. The results show that the re-optimized trajectory (#82875) remains stable under similar error conditions. This improvement demonstrates the model's ability to learn robustness from real errors and proves that data-driven optimization can actually improve the fault tolerance of the device operation under safety constraints.

Experimental Results Proving Improved Robustness

Another experiment, the "predict-first" extrapolation test, verifies the model's generalization potential in unknown parameter ranges. The researchers increased the current upper limit from 140 kA to 170 kA and generated the trajectory solely based on the predictions of the neural state-space model (NSSM) before the experiment. The experimental results show that the model's predictions of key physical quantities are highly consistent with the measurements, and the discharge was successfully terminated without rupture.

Prior Predictions and Experimental Results of the Extrapolation Test Scenario

Advancing the Journey of Turning the "Most Ideal Energy" into Reality

It is reported that the research team is collaborating with the Commonwealth Fusion Systems (CFS) to study how to use the new prediction model and similar tools to better predict plasma behavior, avoid machine disruptions, and achieve safe fusion power generation. Team member Allen Wang said, "We are working hard to solve scientific problems to achieve the regular application of nuclear fusion. Although what we're doing now is just the beginning of a long journey, I think we've made some good progress." In addition, many new studies have emerged in this interdisciplinary field.

The Princeton Plasma Physics Laboratory (PPPL) in the United States, in collaboration with several universities, proposed Diag2Diag. This model can virtually reconstruct key plasma parameters when some sensors fail or observations are limited by learning the correlation between multi-source diagnostic signals, significantly improving the monitoring and early warning capabilities of fusion devices. The relevant research, titled "Diag2Diag: AI-enabled virtual diagnostics for fusion plasma," was published on the arXiv platform. Paper Link: https://arxiv.org/abs/2405.05908v2

In addition, a study titled "FusionMAE: large-scale pretrained model to optimize and simplify diagnostic and control of fusion plasma" published on the arXiv platform proposed a large-scale self-supervised pre-trained model, FusionMAE, for fusion control systems. This model integrates more than 80 diagnostic signals into a unified embedding space and learns the potential correlations between different channels through a masked autoencoder (MAE) structure, achieving efficient alignment of diagnostic and control data streams and pioneering the integration of large-scale artificial intelligence models in the field of fusion energy. Paper Link: https://arxiv.org/abs/2509.12945

Undoubtedly, artificial intelligence is becoming an indispensable force in the journey of turning the "most ideal energy - nuclear fusion energy" into reality.

References:

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

This article is from the WeChat official account "HyperAI Super Neural". Author: Paida Xingxing, Editor: Li Baozhu. Republished by 36Kr with permission.