Abstract
Simulation-based plasma scenario development, optimization, and control are crucial elements toward the successful deployment of next-generation experimental tokamaks and fusion power plants. Current simulation codes require extremely intensive use of HPC resources that make them unsuitable for iterative or real-time applications. Neural network-based surrogate models of expensive simulators have been proposed to speed up such costly workflows. In this work, neural PDE surrogates are trained for the JOREK MHD code and the STORM scrape-off layer code using the PDEArena library. The performance of these surrogate models is investigated as a function of training set size as well as for long-term predictions. It is found that surrogates trained on more data perform best for both long- and short-term predictions. Additionally, surrogate models trained on multiple variables achieve higher accuracy and more stable performance. Downsampling the training set in time may provide stability in the long term at the expense of short-term predictive capability, though visual inspection suggests that multiple metrics should be used to evaluate performance.
Introduction
Developing fusion energy requires precise control over plasma—the ultra-hot, turbulent "fourth state of matter." To design and operate future reactors like ITER or STEP, scientists rely on high-fidelity simulation codes such as JOREK for magnetohydrodynamics and STORM for the plasma edge. However, these simulations are computationally expensive, often requiring massive High-Performance Computing (HPC) clusters and significant time to run. This latency makes them unsuitable for real-time reactor control or rapid design optimization.
AI-based surrogate models offer a promising solution by providing fast approximations of these complex physics. While previous efforts in the fusion community were often limited to simple 1D models or point estimates, this paper investigates 2D surrogate models that are data-efficient and capable of stable, long-term predictions.
Methodology
The researchers utilized Fourier Neural Operators (FNOs) as the core architecture. Unlike standard Convolutional Neural Networks (CNNs), which are tied to a specific pixel grid, FNOs learn mappings between function spaces in the Fourier domain. This makes them "discretization-invariant," allowing a model trained on one resolution to potentially generalize to others.
To handle temporal evolution, the study employed autoregressive rollouts. In this setup, the model takes a window of previous time steps as input to predict the next slice, which is then iteratively fed back into the model to generate a long-term sequence.
Key methodological innovations included:
- Multivariable Training: Training single models to learn multiple coupled physical variables (e.g., density and electric potential) simultaneously.
- Time Downsampling: Experimenting with sparser training intervals to improve long-term stability.
- PDEArena Integration: Leveraging Microsoft’s
PDEArenalibrary for standardized benchmarking of neural PDE solvers.
Datasets
The study evaluated performance on two primary datasets:
- JOREK Dataset: 2,000 simulations of multiple plasma blobs with non-uniform temperature in a 2D slab geometry, simulating large-scale instabilities.
- STORM Dataset: 1,000 simulations of plasma turbulence in the scrape-off layer (the plasma edge), involving complex fluid dynamics and electric potentials.
Results and Discussion
The Power of Multivariable Learning
A key finding of the study is that multivariable models significantly outperform single-variable ones. Because plasma variables are physically coupled, a model that observes multiple variables simultaneously can better capture the underlying physics. This approach prevents the "divergence"—where predictions become physically nonsensical—often seen when variables are modeled in isolation.
Data Efficiency and Stability
As expected, increasing the training set size (e.g., from 300 to 1,800 trajectories) consistently improved performance across both datasets. However, the study also highlighted a trade-off regarding temporal resolution: downsampling training data in time can act as a regularizer, improving stability for very long-term predictions at the cost of capturing fine-grained, short-term details.
Metrics vs. Visual Fidelity
While the FNOs achieved impressive test Mean Squared Errors (MSE)—approximately 3.14e-4 for JOREK and 2.02e-4 for STORM—the authors noted that MSE alone can be misleading. Standard error metrics often favor "blurry" results that capture low-frequency features but miss high-frequency turbulence. Visual inspection remains a critical component of evaluating whether a surrogate model preserves the essential physics of the simulation.
Conclusion
This work demonstrates that FNO-based surrogate models can achieve orders-of-magnitude speedups over traditional HPC codes while maintaining high accuracy. Future research will focus on extending these models to 3D toroidal geometries characteristic of real tokamaks and exploring transfer learning techniques to bridge the gap between low-fidelity and high-fidelity simulation data.
Summary updated